Download the full version of the ebook now at ebookultra.

com

Generating random networks and graphs First

Edition Ton Coolen

https://ebookultra.com/download/generating-random-
networks-and-graphs-first-edition-ton-coolen/

Explore and download more ebook at https://ebookultra.com

 Recommended digital products (PDF, EPUB, MOBI) that
you can download immediately if you are interested.

Introduction to Random Graphs Alan Frieze

https://ebookultra.com/download/introduction-to-random-graphs-alan-
frieze/

ebookultra.com

Probability on Graphs Random Processes on Graphs and

Lattices Institute of Mathematical Statistics Textbooks
Geoffrey Grimmett
https://ebookultra.com/download/probability-on-graphs-random-
processes-on-graphs-and-lattices-institute-of-mathematical-statistics-
textbooks-geoffrey-grimmett/
ebookultra.com

Projected costs of generating electricity Oecd

https://ebookultra.com/download/projected-costs-of-generating-
electricity-oecd/

ebookultra.com

Field Guide to Probability Random Processes and Random

Data Analysis Larry C. Andrews

https://ebookultra.com/download/field-guide-to-probability-random-
processes-and-random-data-analysis-larry-c-andrews/

ebookultra.com
Probability and Random Variables G.P. Beaumont

https://ebookultra.com/download/probability-and-random-variables-g-p-
beaumont/

ebookultra.com

Experiments in Holism Theory and Practice in Contemporary

Anthropology 1st Edition Ton Otto

https://ebookultra.com/download/experiments-in-holism-theory-and-
practice-in-contemporary-anthropology-1st-edition-ton-otto/

ebookultra.com

Spectral Radius of Graphs 1st Edition Dragan Stevanovic

https://ebookultra.com/download/spectral-radius-of-graphs-1st-edition-
dragan-stevanovic/

ebookultra.com

Algebraic Elements of Graphs 1st Edition Yanpei Liu

https://ebookultra.com/download/algebraic-elements-of-graphs-1st-
edition-yanpei-liu/

ebookultra.com

Spectral Geometry of Graphs 1st Edition Pavel Kurasov

https://ebookultra.com/download/spectral-geometry-of-graphs-1st-
edition-pavel-kurasov/

ebookultra.com
Generating random networks and graphs First Edition
Ton Coolen Digital Instant Download
Author(s): Ton Coolen, Alessia Annibale, Ekaterina Roberts
ISBN(s): 9780198709893, 0198709897
Edition: First edition
File Details: PDF, 16.54 MB
Year: 2017
Language: english
GENERATING RANDOM NETWORKS AND GRAPHS
Generating Random Networks and Graphs

A.C.C. Coolen
Department of Mathematics, King’s College London, UK

A. Annibale
Department of Mathematics, King’s College London, UK

E.S. Roberts
Department of Mathematics, King’s College London, UK

3
 3
Great Clarendon Street, Oxford, OX2 6DP,
United Kingdom
Oxford University Press is a department of the University of Oxford.
It furthers the University’s objective of excellence in research, scholarship,
and education by publishing worldwide. Oxford is a registered trade mark of
Oxford University Press in the UK and in certain other countries
© A.C.C. Coolen, A. Annibale, and E.S. Roberts 2017
The moral rights of the authors have been asserted
First Edition published in 2017
Impression: 1
All rights reserved. No part of this publication may be reproduced, stored in
a retrieval system, or transmitted, in any form or by any means, without the
prior permission in writing of Oxford University Press, or as expressly permitted
by law, by licence or under terms agreed with the appropriate reprographics
rights organization. Enquiries concerning reproduction outside the scope of the
above should be sent to the Rights Department, Oxford University Press, at the
address above
You must not circulate this work in any other form
and you must impose this same condition on any acquirer
Published in the United States of America by Oxford University Press
198 Madison Avenue, New York, NY 10016, United States of America
British Library Cataloguing in Publication Data
Data available
Library of Congress Control Number: 2016946811
ISBN 978–0–19–870989–3
Printed and bound by
CPI Group (UK) Ltd, Croydon, CR0 4YY
Links to third party websites are provided by Oxford in good faith and
for information only. Oxford disclaims any responsibility for the materials
contained in any third party website referenced in this work.
Preface

Networks have become popular tools for characterizing and visualizing complex sys-
tems with many interacting variables. In the study of such systems, it is often helpful
to be able to generate random networks with statistical features that mimic those of
the true network at hand. These random networks can serve as ‘null hypotheses’ in
hypothesis testing, or as proxies in simulation studies. How to generate random net-
works with controlled topological characteristics accurately and efficiently has become
an important practical challenge, and is in itself an interesting theoretical question.
Random networks result, by definition, from stochastic processes. The natural lan-
guage for graph generation protocols is therefore that of network or graph ensembles.
These ensembles are constrained by the statistical features that one chooses to im-
pose. The core of the network generation problem is to construct stochastic processes
or algorithms that, upon equilibration, sample the space of all networks that have
the desired features, but are otherwise unbiased. Moreover, these algorithms should be
able to generate large networks within practical time scales.
We set out to write a concise but self-contained text on the generation of random
networks with controlled topological properties, including the mathematical deriva-
tions of relevant algorithms and their numerical implementations. Such details appear
to be under-represented in the presently available books. They tend to be found in
discipline-specific journals that are not easily accessible, and research articles devoted
to one algorithm at a time. We have tried to bring together the most important results
that are scattered over the physics, combinatorial mathematics and computer science
literature, and explain with uniform notation the rationale, the potential and limita-
tions, and application domains of each approach. We seek to explain the theoretical
issues and the practical algorithms resulting from analysis of the problem in such a
way that the reader is able to focus on either aspect if he or she so wishes. To achieve
this, we have separated (where possible) practical recipes from theoretical arguments,
to aid selective navigation driven by personal and domain-specific preferences and
needs. More extensive mathematical arguments are delegated to appendices, and we
have supplemented the main text with examples and exercises with solutions.
This book is aimed at established or junior researchers and their graduate or ad-
vanced undergraduate students in computer science, quantitative biology, social sci-
ence, ecology, bioinformatics, applied mathematics and theoretical physics, whose work
involves networks or graphs. We hope that it may serve as a useful reference, as a text-
book for postgraduate lecture courses and as an inspiration for further quantitative
research.

London, November 2016 Ton Coolen, Alessia Annibale and Kate Roberts
Acknowledgements
It is our great pleasure to thank the many colleagues and students with whom over the
years we have enjoyed discussing the modelling and generation of graphs and networks.
In particular, we would like to mention (in alphabetical order) Elena Agliari, Ginestra
Bianconi, Zdislaw Burda, Sun Sook Chung, Owen Courtney, Andrea De Martino,
Luis Fernandes, Thomas Fink, Franca Fraternali, Clara Grácio, Alexander Hartmann,
Jens Kleinjung, Reimer Kühn, Alexander Mozeika, Alessandro Pandini, Nuria Planell-
Morell, Conrad Perez-Vicente and Peter Sollich.
Contents

PART I THE BASICS

1 Introduction 3
2 Definitions and concepts 9
2.1 Definitions of graphs and their local characteristics 9
2.2 Macroscopic characterizations of graphs 11
2.3 Solutions of exercises 17
3 Random graph ensembles 19
3.1 The Erdös–Rényi graph ensemble 20
3.2 Graph ensembles with hard or soft topological constraints 22
3.3 The link between ensembles and algorithms 25
3.4 Solutions of exercises 28

PART II RANDOM GRAPH ENSEMBLES

4 Soft constraints: exponential random graph models 33
4.1 Definitions and basic properties of ERGMs 33
4.2 ERGMs that can be solved exactly 36
4.3 ERGMs with phase transitions: the two-star model 40
4.4 ERGMs with phase transitions: the Strauss (triangle) model 47
4.5 Stochastic block models for graphs with community structure 57
4.6 Strengths and weaknesses of ERGMs as null models 59
4.7 Solutions of exercises 62
5 Ensembles with hard constraints 64
5.1 Basic properties and tools 64
5.2 Nondirected graphs with hard-constrained number of links 68
5.3 Prescribed degree statistics and hard-constrained number of links 70
5.4 Ensembles with prescribed numbers of links and two-stars 73
5.5 Ensembles with constrained degrees and short loops 75
5.6 Solutions of exercises 79

PART III GENERATING GRAPHS FROM GRAPH ENSEMBLES

6 Markov Chain Monte Carlo sampling of graphs 83
6.1 The Markov Chain Monte Carlo sampling method 83
6.2 MCMC sampling for exponential random graph models 89
6.3 MCMC sampling for graph ensembles with hard constraints 94
6.4 Properties of move families – hinge flips and edge swaps 97
6.5 Solutions of exercises 102
viii Contents

7 Graphs with hard constraints: further applications and ex-

tensions 105
7.1 Uniform versus non-uniform sampling of candidate moves 105
7.2 Graphs with a prescribed degree distribution and number of links 112
7.3 Ensembles with controlled degrees and degree correlations 117
7.4 Generating graphs with prescribed degrees and correlations 121
7.5 Edge swaps revisited 128
7.6 Non-uniform sampling of allowed edge swaps 131
7.7 Non-uniform sampling from a restricted set of moves: a hybrid
MCMC algorithm 135
7.8 Extensions to directed graphs 139
7.9 Solutions of exercises 141

PART IV GRAPHS DEFINED BY ALGORITHMS

8 Network growth algorithms 153
8.1 Configuration model 153
8.2 Preferential attachment and scale-free networks 157
8.3 Analyzing growth algorithms 163
8.4 Solutions of exercises 166
9 Specific constructions 169
9.1 Watts–Strogatz model and the ‘small world’ property 169
9.2 Geometric graphs 173
9.3 Planar graphs 177
9.4 Weighted graphs 179
9.5 Solutions of exercises 181

PART V FURTHER TOPICS

10 Graphs on structured spaces 187
10.1 Temporal networks 187
10.2 Multiplex graphs 193
10.3 Networks with labelled nodes 197
10.4 Relations and connections between models 202
10.5 Solutions of exercises 204
11 Applications of random graphs 207
11.1 Power grids 207
11.2 Social networks 208
11.3 Food webs 212
11.4 World Wide Web 215
11.5 Protein–protein interaction networks 216
Key symbols and terminology 221
Appendix A The delta distribution 229
Appendix B Clustering and correlations in Erdös–Rényi graphs 231
 Contents ix

B.1 Clustering coefficients 231

B.2 Degree correlations in the sparse regime 233
Appendix C Solution of the two-star exponential random graph
model (ERGM) in the sparse regime 236
Appendix D Steepest descent integration 240
Appendix E Number of sparse graphs with prescribed average
degree and degree variance 242
Appendix F Evolution of graph mobilities for edge-swap dy-
namics 244
Appendix G Number of graphs with prescribed degree sequence 247
Appendix H Degree correlations in ensembles with constrained
degrees 249
Appendix I Evolution of triangle and square counters due to
ordered edge swaps 253
Appendix J Algorithms 256
J.1 Algorithms based on link flips 259
J.2 MCMC algorithms sampling from graphs with hard constraints 268
J.3 Algorithms based on hinge flips 271
J.4 Algorithms based on edge swaps 277
J.5 Growth Algorithms 282
J.6 Auxiliary routines 289
References 297
Index 309
Part I
The basics
1
Introduction

Networks and their popularity. Our ability to collect and store vast amounts of quan-
titative data in many areas of the natural and the man-made world has increased
dramatically. These include data of a socio-economic nature, such as social links be-
tween individuals or professional collaboration networks, consumer preferences and
commercial interactions, trade dependencies among corporations and credit or insur-
ance obligations between financial institutions. They include usage and traffic data
on computer and telephone communication systems, satellite networks, the Internet,
electricity grids, rail, road or air travel connections and distribution networks. We
also collect and process large amounts of geological and meteorological data, data on
sea levels, air and water pollution, volcanic and seismic records, and sizes of polar
and glacier ice sheets. Last but not least, in recent years we have seen an explosion
of biomedical information, such as experimental data on biochemical processes and
structures at cellular, sub-cellular and even molecular levels, the topologies of com-
plex composite organs such as the brain or the immune system, genomic and epigenetic
data (gene expression levels, DNA mutations), epidemiological data and vast numbers
of patient records with clinical information.
However, one tends to collect data for a reason. This reason is in many cases the
desire to understand the dynamical behaviour of the system under study to such an
extent that one can predict its future evolution, and its response to perturbations or
interventions. We may want this for academic reasons, for financial gain, to improve
and optimize a system’s efficiency, to design effective regulatory controls or (in the
case of medicine) to understand and cure diseases. For small and simple systems,
the translation of observed data into qualitative and quantitative understanding of
function is usually not difficult. We do not have to be engineers to understand the
workings of a bicycle once we have seen it in action, and after we have made an
inventory of the parts and their links. In contrast, if we collect data on complex
systems with millions or more of nonlinearly interacting variables, just having a list
of parts and their connections and/or many observations of their collective patterns
of behaviour is no longer enough to understand how these systems work. This is even
more true if we do not know the detailed forces via which the components interact.
A first and often useful step in modelling and visualizing structural data on com-
plex many-variable systems is to represent them as networks or graphs1 . The idea is
to represent each observed component as a node in the network, and each observed in-
teraction between two components as a link. Dependent upon the application domain,

1 In this book we will use the words ‘network’ and ‘graph’ interchangeably.

Generating Random Networks and Graphs. First Edition. A.C.C. Coolen, A. Annibale, and
E.S. Roberts. c A.C.C. Coolen, A. Annibale, and E.S. Roberts 2017. Published in 2017 by
Oxford University Press. DOI: 10.1093/acprof:oso/9780198709893.001.0001
4 Introduction

the nodes of such networks may then represent anything ranging from genes, molecules
or proteins (in biology), via processors or servers (in computer science), to people, air-
ports, power plants or corporations (in social science or economics). The links could
refer to biochemical reactions, wired or wireless communication channels, friendships,
financial contracts, etc. The result is an appealing and intuitive representation of the
system, but the price paid is a significant loss of information and ambition. Limiting
ourselves to the network representation means giving up a lot of information; we only
record which parts interact, and disregard how and when they interact. The rationale
is that much can perhaps already be learned from the topologies of the networks alone.
For example, DNA contains the assembly instructions for complicated macro-
molecules called proteins. Proteins are the universal workhorses of the cell. They serve
as building material, molecular factories, information readers, translators, transport
vehicles, sensors and messengers. They interact with each other by forming complexes,
which are semi-stable supermolecules formed of multiple proteins that attach to each
other selectively. The network approach to this system would be to record – from the
results of many complicated biochemical protein binding experiments – only which
protein species are present (to be represented by network nodes) and which of these
species can interact (to be represented by nondirected links between the nodes of in-
teracting species, i.e. links without arrows); see Figure 1.1 (based on data from [149]).
Typical databases contain some 70,000–100,000 reported interactions – still only about
half of the interactions believed to exist in the human proteome – and the problem
for the modeller is how to visualize and utilize all this information. Similarly, in the
analysis of financial markets, we might represent each financial institution or large
market player (banks, hedge funds, insurers, pension funds, regulators, large commod-
ity traders, central banks, market makers, etc.) as a network node, and connect those
nodes that are known to have mutual contractual or regulatory obligations, or that
have a trading relationship. In the analysis of ecosystems, we could represent each
plant or animal species by a node, and connect those nodes that have either a direct
dependency in the food chain, or a symbiotic relationship or that interact indirectly
by creating environmental conditions that affect each other’s survival likelihood. In
all those cases, one would then proceed to analyze topological features of the created
networks, and try to find whether these correlate with important functional character-
istics of the complex systems they are meant to represent. In these last two examples,
we would use directed networks (i.e. with links that carry arrows), since the mutual
influences need not be reciprocal. The relevant characteristics to be correlated with
network topology will vary from system to system, and will also be dependent upon
the interests of the modeller. In financial markets and ecosystems, the main interests
would probably be in understanding the origin and nature of anomalous fluctuations,
the topological factors that influence overall stability and large crashes or wipeouts,
and in determining whether and how the dynamics of these systems can be predicted
and controlled. In large distribution networks of, e.g. food, electricity, fuel, people or
information, one would be interested in questions relating to path lengths (since these
determine transport costs), vulnerability against spontaneous of maliciously induced
breakdowns, and how to design the networks to minimize this vulnerability.
 Introduction 5

Fig. 1.1 Bottom: Extract of raw data on human protein–protein interactions as reported
in biological databases, being lists of pairs of protein species that were observed to interact,
together with their formal codes and information on the experiments where their interaction
was observed. Each line reports on one pair interaction. Top: The network representation of
the data. Each protein is represented by a node, and each pair interaction, i.e. each line in
the table, is represented by a link that connects the two nodes concerned.
6 Introduction

Network literature. There are many excellent books on the analysis and applications
of networks and graphs. Some touch briefly upon certain aspects of graph generation,
usually by mentioning the generation of Erdös–Rényi graphs, and selected existing
algorithms that would produce specific classes of graphs (often without derivations).
Here we limit ourselves to mentioning a few popular books, and apologize to the
authors of those that were left out. For example, [26] is a new edition of an influential
classic, which covers mainly the formal mathematical theory of random graphs. The
volume [63] is a popular text, somewhat less mathematical, but with many real-world
examples of networks. A large collection of chapters written by different experts on
networks and their applications is found in [30]. The monograph [65] deals mainly with
the mathematical analysis and application of random graphs processes on graphs. The
most comprehensive text on networks presently available is probably [131]. Other
recommended books are [5], [23] and [40].

The role of random networks. With larger and larger datasets, it becomes crucial to
have computationally practical and analytically well-understood algorithms for gener-
ating networks. This is especially so because the ability to test a hypothesis against a
properly specified control case (a null model ) is at the heart of the scientific method.
Hence, knowledge on how to generate samples from unbiased and well-controlled ran-
dom graph ensembles is vital for the maturing field of network science. The null model
approach is appealing in its conceptual simplicity. It effectively provides synthetic
‘data’, which can be analyzed in the same way as the real dataset. One can then
learn which observed properties are particular to the real dataset (i.e. ‘special’ and
informative), and which are common within the wider ensemble of which the observed
network and the null model are members. Applications of network null models are
consequently wide-ranging and central to network science. For instance, [164] applies
null models to identify over-represented motifs in the transcriptional regulation net-
work of E. coli; [180] discusses adapting the Watts–Strogatz method to generating
random networks to model power grids; [136] explores motifs found within an inter-
firm network; [132] uses network null models to study social networks; [115] compares
topological properties of interaction and transcription regulatory networks in yeast
with randomized null model networks and postulates that links between highly con-
nected proteins are suppressed in protein interaction networks; finally [81] discusses
the challenges of specifying a suitable matrix null model in the field of ecology.

Non-trivial aspects of graph generation. The key characteristics of network generation

algorithms are their flexibility (e.g. the extent to which the topological features can
be controlled), their uniformity (i.e. the extent to which they generate graphs with
the correct probabilities), and their computational efficiency. These three are usually
difficult to optimize simultaneously. The most commonly used methods for generating
random networks with a given degree sequence (i.e. the full specification of the num-
ber of interaction partners of each node) are variations on the so-called stub-joining
and edge-swap algorithms. The former is based on ‘growing’ links, starting from a
set of disconnected nodes. The edge-swap method starts with a network with the de-
sired degree sequence, and then randomizes it by repeatedly swapping pairs of edges.
In both algorithms, the main problem lies in accurately controlling the probabilities
 Introduction 7

with which graphs are generated, while keeping computational demands within fea-
sible limits. Another class of models is that of exponential random graph ensembles.
Here the graph probabilities are from the start written in the so-called Boltzmann
form, with a specially constructed energy function assigned to each possible graph
realization. These are particularly popular in the social sciences, and can be gener-
ated by the Metropolis–Hastings algorithm, where edges are added or removed at
random. For such networks the probabilities are controlled trivially, although at high
computational cost, but building in ‘hard’ structural constraints is a difficult problem.
Finally, many models have been proposed which are specifically designed to mimic real-
world networks of a certain type, of which Barabási’s preferential attachment network
growth model is a well-known example. Systems biology uses graph randomization
protocols to identify the elementary bricks (motifs, i.e. small subgraphs that may be
over-represented) characterizing cellular networks [164, 93]. Many other examples can
be found in economics, ecology and the social sciences; see, e.g. [165, 166]. In all cases,
it is crucial that the synthetic networks generated as null models are representative
of the underlying ensembles. Any inadvertent bias in the network generation process
may invalidate the hypothesis test. Generating such bias-free networks is known to be
a non-trivial problem, that has produced much inspiring work and some hard open
questions [17, 128, 131]. Using biased algorithms for producing null models is funda-
mentally unsound, and unacceptable when there are rigorous unbiased alternatives.
When more complicated observables than the degree sequence are involved (like de-
gree correlations, clustering or loop statistics), the situation rapidly becomes more
difficult, both mathematically and numerically.

Overview of this book. An algorithm is a dynamical process. An algorithm that gen-

erates a random graph is by construction a stochastic process. Sometimes one defines
random graphs by a desired probability distribution, and one needs to find a corre-
sponding process that leads to this distribution, which can then serve as an algorithm.
Alternatively, dependent upon one’s problem domain, it could be more natural to de-
fine random graphs directly in terms of a process; one then needs to understand the
probability distribution to which this process evolves in order to predict or have con-
trol over the end result. Although it may sometimes seem when reading the literature
that these two approaches to random graphs (and their user groups) are quite dis-
tinct, it should be clear that they simply represent different entry routes into the same
arena. At the core of the problem always lies the need to establish the mathematical
connection between stochastic processes for graphs (usually taking the form of Markov
chains) and the stationary probability distributions to which these processes evolve.
This unified view of the graph generation problem allows us to bring most ap-
proaches to graph generation under the same conceptual and notational umbrella, and
is reflected in the structure of this book. In Part 1, we start with notation definitions,
and introduce basic concepts in quantifying network topologies and in describing tai-
lored families of random graphs probabilistically. This material will be largely familiar
to those who have worked on network theory before. In Part 2, we discuss random
graphs that are specified by a desired probability distribution, with soft or hard struc-
tural constraints, in the language of so-called graph ensembles. In Part 3, we discuss
8 Introduction

the construction Markov chain algorithms that upon equilibration will generate graphs
from specified graph ensembles. In Part 4, we discuss those cases where one starts from
the opposite side, i.e. from a stochastic process specification (i.e. an algorithm), and
where one seeks to understand and predict the stationary probabilities to which this
process will evolve. Part 5 is devoted to further topics and applications of random
graphs. Most appendices deal with technical issues which, although important, in-
terrupt the flow of the book. The final appendix provides a collection of algorithms
derived and discussed in the main text.
2
Definitions and concepts

2.1 Definitions of graphs and their local characteristics

Overview: In this section, we will define and explain the following concepts and as-
sociated notation: nodes, links, degree, adjacency matrix and clustering coefficient.

Notation, nodes and links. We denote by N the number of nodes (or vertices) in a
graph, and we label these N nodes with lower-case Roman indices i, j, k ∈ {1, . . . , N }.
For each node pair (i, j) we next define Aij = 1 if a link (or edge) j → i is present
in the graph, and Aij = 0 if it is not. The set of N 2 variables Aij ∈ {0, 1} thus
specifies the graph in full, and is abbreviated as A. One can also interpret the binary
numbers {Aij } as the entries of an N ×N matrix, which is called the graph’s adjacency
matrix. Nondirected graphs are those in which Aij = Aji for all (i, j), so the links
in such graphs carry no arrows. The number ofPlinks L in a directed graph, where
node pairs exist with Aij 6= Aji , equals L = i,j≤N Aij . In a nondirected graph,
we do P not count AijP= 1 and Aji = 1 separately, so here the number of links is
L = 12 i6=j≤N Aij + i≤N Aii . In this book, we will usually limit ourselves to graphs
without self-interacting nodes, so Aii = 0 for all i. These are called simple graphs.

Microscopic structure characterization in terms of local features. To characterize graph

topologies more intuitively, we next inspect simple quantities that reveal aspects of the
structure of the graph in the vicinity of individual nodes. The first of these quantities
are the in- and out-degrees kiout (A) ∈ IN and kiout (A) ∈ IN of each node i in graph
A. They count, respectively, the number of arrows flowing into and out of a node i
N
X N
X
kiin (A) = Aij , kiout (A) = Aji (2.1)
j=1 j=1

We will denote the pair of degrees for a node i as ~ki (A) = (kiin (A), kiout (A)) ∈ IN2 .
We see that
PN the total number of links
PN in a directed graph can now be written as either
L(A) = i=1 kiin (A) or L(A) = i=1 kiout (A). The average in-degree and the average
out-degree in a graph are always identical, which reflects the simple fact that all arrows
flowing out of a node will inevitably flow into another node. In nondirected graphs,
where Aij = Aji for all (i, j), we find that kiin (A) = kiout (A)
P for all i. Here we can drop
the superscripts and simply refer to the degree ki (A) = j Aij ∈ IN of each node i.
In nondirected graphs without self-interactions, the number of links is then seen to be
L(A) = 12 N k̄(A), where k̄(A) = N −1 i≤N ki (A) is the average degree in the graph.
P

Generating Random Networks and Graphs. First Edition. A.C.C. Coolen, A. Annibale, and
E.S. Roberts. c A.C.C. Coolen, A. Annibale, and E.S. Roberts 2017. Published in 2017 by
Oxford University Press. DOI: 10.1093/acprof:oso/9780198709893.001.0001
10 Definitions and concepts

AA
K A A
A A
• •
A
•@ A• •@ A•
@
I
@ 6 @ @
@ @ @
@• -• - @• •A @• •A

i A
AA i A i A
U A A
• • •
kiin (A) = 1 ki (A) = 4 ki (A) = 4
kiout (A) =3 Ci (A) = 1/6 Ci (A) = 2/6 = 1/3

Fig. 2.1 Left: Illustration for directed graphs of the meaning of in- and out-degrees kiin (A)
and kiout (A), i.e. the number of arrows flowing into and out of the given node. Middle and
right: Illustration for nondirected graphs of degrees ki (A) and clustering coefficients Ci (A).
The latter gives the number of distinct neighbour pairs of i that are themselves connected,
divided by the total number of such pairs. Due to the absence of link directionality, there is
no distinction in nondirected graphs between in- and out-degrees.

There are several ways to characterize a graph’s local structure beyond the level of
counting links from a node i to its immediate neighbours. For instance, for nondirected
graphs without self-links, one defines the clustering coefficient Ci (A) ∈ [0, 1], which
gives the fraction of nodes linked to i that are themselves mutually connected
number of connected node pairs among neighbours of i
Ci (A) =
number of node pairs among neighbours of i
PN
j,k=1 (1−δjk )Aij Ajk Aik 2Ti (A)
= PN = (2.2)
j,k=1 (1−δjk )Aij Aik
k i (A)[k i (A) − 1]

PN
Here Ti (A) = 12 j,k=1 Aij Ajk Aki ∈ IN counts the number of distinct triangles, i.e. the
number of nondirected loops of length three in the graph in which node i participates
[163]. The factor 12 in Ti (A) corrects for overcounting: any nondirected triangle starting
from node i can be drawn with two possible orientations. We also used Kronecker’s
shorthand δab , defined as
 Z π
1 if a = b dω iω(a−b)
δab = δab = e (2.3)
0 if a 6= b −π 2π

The mathematically equivalent integral formula for δab on the right will turn out to
be a helpful tool for simplifying various calculations. Clearly, the clustering coefficient
can only be defined for nodes i with degree ki (A) ≥ 2. In Fig. 2.1, we illustrate the
meaning of degrees and clustering coefficients with some simple examples.
The above quantities can be generalized in obvious ways. For instance, the gener-
alized degrees count the number of distinct paths of a given length ` ≥ 1 that either
flow out of, or into a node i:
 Macroscopic characterizations of graphs 11

N N
(`)in
X X
ki (A) = ... Aij1 Aj1 j2 . . . Aj`−1 j` (2.4)
j1 =1 j` =1
N N
(`)out
X X
ki (A) = ... Aj` j`−1 . . . Aj2 j1 Aj1 i (2.5)
j1 =1 j` =1

For ` = 1, these formulae reduce to the previous expressions (2.1). We can similarly
introduce quantities that generalize the triangle counters Ti (A) by counting closed
paths in the graph A of arbitrary lengths ` ≥ 3 that pass through node i.

2.2 Macroscopic characterizations of graphs

Overview: In this section, we define and explain the following concepts and associated
notation: degree distribution, degree–degree correlation and adjacency matrix spectra.

Macroscopic characterization based on statistics of local features. The quantities de-

fined so far provide microscopic structural information in the vicinity of individual
nodes. We can characterize the structure of the full graph by listing for all N nodes
the values of the locally defined quantities. In nondirected graphs, we would then ob-
tain, for instance, the degree sequence k(A) = (k1 (A), . . . , kN (A)) ∈ INN . In directed
graphs, where each node has two degrees ~ki (A) = (kiin (A), kiout (A)), this degree se-
quence would take the form k(A) = (~k1 (A), . . . , ~kN (A)) ∈ IN2N . We can make similar
lists of quantities such as the clustering coefficients.
For large graphs, or when comparing graphs of different sizes, it will be more helpful
to define quantities that are intrinsically macroscopic in nature. The simplest of these
are histograms of the observed values of the previously defined local features. If we
divide, for each possible value, how often this value is observed by the total number
of observations (i.e. the number of nodes), we obtain the empirical distribution of the
given feature in the graph. For example, we can define the degree distribution of a
nondirected graph via

N
1 X
∀k ∈ IN : p(k|A) = δk,ki (A) (2.6)
N i=1

The quantity p(k|A) represents the probability that a randomly drawn node in the
nondirected graph A will have degree k. Similarly we can define the joint distribution
p(k, T |A) of degrees and triangle numbers

N
1 X
∀k, T ∈ IN : p(k, T |A) = δk,ki (A) δT,Ti (A) (2.7)
N i=1

Now p(k, T |A) is the probability that a randomly drawn node will have degree k and
participate in T triangles. For directed graphs, the equivalent definition is
12 Definitions and concepts

N
1 X
∀(k in , kout ) ∈ IN2 : p(k in, kout |A) = δ in in δ out out (2.8)
N i=1 k ,ki (A) k ,ki (A)

Here p(k in, kout |A) represents the probability that a randomly drawn node in the graph
will have the pair (k in, kout ) of in- and out-degrees.

Exercise 2.1 Calculate the average degree hki = k≥0 p(k)k and the degree variance σk2 =
P

hk2 i − hki2 for the degree distribution p(k) = e−γk /CN for k ∈ IN, with γ > 0. Next, consider
the following scale-free distribution: p(k) = 0 for k = 0 or k > N , and p(k) = k−γ /CN for
0 < k ≤ N , again with γ > 0. Give a formula for CN . For which γ is p(k) normalizable as
N → ∞? Give formulae for the average degree and the degree variance. For which γ values
are the average degree and the variance finite as N → ∞? Give an estimate of Rthe average
N
and variance for γ = 2.5 and N = 10,000, using the approximation N −λ
≈ 1 dk k −λ .
P
k=1 k

Macroscopic characterization involving degree correlations. The interpretation of de-

gree distributions in terms of drawing nodes in the graph at random suggests additional
macroscopic routes for characterizing graph structures. For instance, for nondirected
graphs A without self-interactions, we can define the probability W (k, k0 |A) that a
randomly drawn link will connect nodes with specified degrees k and k 0 as
PN
0 i,j=1 δk,ki (A) Aij δk0 ,kj (A)
W (k, k |A) = PN
i,j=1 Aij
N
1 X
= δk,ki (A) Aij δk0 ,kj (A) (2.9)
N k̄(A) i,j=1

Clearly W (k, k 0 |A) = W (k 0, k|A) for all (k, k0 ) ∈ IN2 , and W (k, k0 |A) = 0 if k = 0 or
k 0 = 0 (or both). From (2.9) follows also the assortativity a(a) [129], which is defined
as the overall Pearson correlation between the degrees of connected node pairs:

2
W (k, k0 |A)kk 0 −
P P
k,k0 >0 k>0 W (k|A)k
a(A) = P P
2 ∈ [−1, 1] (2.10)
2
k>0 W (k|A)k − k>0 W (k|A)k

with the marginal distribution W (k|A) = k0 >0 W (k, k 0 |A). If a(A) > 0, there is a
P
preference in the graph for linking high-degree nodes to high-degree nodes and low-
degree nodes to low-degree nodes; such graphs are called assortative. If a(A) < 0, the
preference is for linking high-degree nodes to low-degree ones; these graphs are called
dissortative. Upon summing the definition (2.9) over k 0 , we see that the marginal
W (k|A) follows directly from the degree distribution:
N
X 1 X
W (k|A) = W (k, k0 |A) = δk,ki (A) ki (A)
k0 >0
N k̄(A) i=1
 Macroscopic characterizations of graphs 13

k
= p(k|A) (2.11)
k̄(A)

The reason why W (k|A) 6= p(k|A) is that in W (k|A) the degree likelihood of nodes is
conditioned on these nodes coming up when picking links at random; this favours nodes
with many links over those with few. In those graphs where there are no correlations be-
tween the degrees of connected nodes, one would find that the joint distribution (2.9) is
simply the product of the respective marginals (2.11), W (k, k0 |A) = W (k|A)W (k 0 |A)
for all k, k0 > 0. Hence, a useful quantity to characterize correlations is the ratio

W (k, k 0 |A) k̄ 2 (A) W (k, k 0 |A)

Π(k, k0 |A) = 0
= (2.12)
W (k|A)W (k |A) kk 0 p(k|A)p(k 0 |A)

This ratio is by definition equal to one for graphs without degree correlations. Any
deviation from Π(k, k0 |A) = 1 will signal the presence of degree correlations.
One can easily convince oneself that the degree correlations captured by Π(k, k 0 |A)
can provide valuable new information that is not contained in the degree distribution
p(k|A). For instance, in Fig. 2.2 we show two networks that have nearly identical
degree distributions, but are nevertheless seen to be profoundly different at the level
of their degree correlations. This shows us that for large graphs there are still many
distinct possible wirings that are consistent with an imposed degree sequence. The
result of calculating the macroscopic characteristics p(k|A) and Π(k, k 0 |A) for the
example protein interaction data of Fig. 1.1 is shown in Fig. 2.3.
For directed networks, the degree correlations are described by a similar function
W (~k, ~k 0 |A), where ~k = (k in, kout ) and ~k 0 = (k 0in, k0out ). This reflects the need in directed
graphs to distinguish between in-in degree correlations, out-out degree correlations and
in-out degree correlations:
PN
i,j=1 δ~ ki (A) Aij δ~
k,~ k0 ,~
kj (A)
~ ~ 0
W (k, k |A) = PN
i,j=1 Aij
N
1 X
= δ~ ~ Aij δ~k0 ,~kj (A) (2.13)
N k̄(A) i,j=1 k,ki (A)

In contrast to the kernel W (k, k0 |A) for nondirected graphs, which was symmetric
under interchanging k and k 0 , here we will generally have W (~k, ~k 0 |A) 6= W (~k 0, ~k|A).

Modularity. Sometimes the prominent structure of a network is modularity, i.e. the

nodes can be divided into groups such that the links preferentially connect node pairs
within the same group. Finding the optimal assignment of a string (x1 , . . . , xN ) of
module labels to the nodes in such a network is a common problem in network appli-
cations. To quantify the extent to which P a simple nondirected graph is modular, one
compares the number Lintra (A) = 12 ij Aij δxi ,xj of intra-modular links in A (the
factor 12 reflects its nondirected nature) to what would have been found if the wiring
had been random. In a random N -node graph A0 withP the same degree sequence
N
(k1 (A), . . . , kN (A)) as A, we would find hLintra (A0 )i = 12 i,j=1 hA0ij iδxi ,xj . Now
14 Definitions and concepts

p(k|A) Π(k, k0 |A)

0.4 40
1.4

0.3 30
1.2
0
k
0.2 20 1

0.8
0.1
10

0.6
0
0 10 20 30
10 20 30 40
k k

0.4 40
1.4

0.3 30
1.2
0
k
0.2 20 1

0.8
0.1
10

0.6
0
0 10 20 30
10 20 30 40
k k

Fig. 2.2 Illustration of the limitations of using only degree statistics to characterize graphs.
The two N = 5000 graphs A shown here (top row versus bottom row) look similar and have
nearly indistinguishable degree distributions p(k|A) (shown as histograms). However, they
differ profoundly at the level of degree correlations, which is visible only after calculating the
functions Π(k, k0 |A), which are shown as greyscale heat maps on the right.
X
hA0ij i = A0ij p(A0ij |k(A)) = p(1|k(A)) (2.14)
A0ij ∈{0,1}

Here p(A0ij |k) is the probability that in a random graph with degree sequence k =
(k1 , . . . , kN ) one finds nodes i and j connected. This must be proportional to ki and
kj , so we estimate that hA0ij i ≈ ki (A)kj (A)/C(A). The value of C(A) is determined by
summing both sides over i and j, giving i,j≤N hA0ij i = [ i≤N ki (A)]2 /C(A). Hence
P P

C(A) = N k̄(A), and we find our estimate hA0ij i = ki (A)kj (A)/N k̄(A). Therefore
N
1 X ki (A)kj (A)
hLintra (A0 )i ≈ δxi ,xj (2.15)
2 i,j=1 N k̄(A)

One now defines the measure of modularity Q(A), apart from an overall scaling factor,
 Macroscopic characterizations of graphs 15

Π(k, k0 |A)
0.4 50 1.4

1.3
40
0.3 1.2

p(k|A) k0 30
1.1

0.2 1

20 0.9

0.1 0.8
10
0.7

0 0.6
0 10 20 30
10 20 30 40 50
k k

Fig. 2.3 The degree distribution p(k|A) (2.6) (left), and the normalized degree correlation
kernel Π(k, k0 |A) (2.12) (right, shown as a greyscale heatmap) for the protein interaction net-
work of Fig. 1.1. Here N ≈ 9000 and k̄(A) ≈ 7.5. Significant deviations from Π(k, k 0 |A) ≈ 1
mark non-trivial structural properties beyond those captured by the degree distribution.

as the difference between the number of intra-modular links in A and the number we
expect to find by accident in a random graph A0 with the same degrees
N 
1 X ki (A)kj (A) 
Q(A) = Aij − δxi ,xj (2.16)
N k̄(A) i,j=1 N k̄(A)

This definition is applicable only if the graph has a non-zero number of links, so k̄(A) >
0. The chosen scaling factor ensures that Q(A) has the convenient size-independent
bounds −1 ≤ Q(A) ≤ 1. Non-modular graphs give Q(A) ≈ 0.

Exercise 2.2 Prove that the modularity (2.16) obeys the general bounds −1 ≤ Q(A) ≤ 1.
Give an example of an N -node simple nondirected graph for which Q(A) = 1.

Other structure characterizations. The properties discussed above have been selected
because they are important network metrics, with a long history in the field of random
graphs. They measure the statistics of local topological properties, where local is either
interpreted from a node-centred point of view, as with p(k|A) and p(~k|A), or from a
link-centred point of view, as with W (k, k0 |A) and W (~k, ~k 0 |A). We will see that it is
indeed generally much easier to incorporate topological characteristics as constraints
into a random graph ensembles if these relate to (distributions of) local properties.
Alternative approaches to quantifying graph topologies involve measures based on
counting typical and minimal path lengths, and the eigenvalue spectrum %(λ|A) of the
adjacency matrix. For nondirected graphs the adjacency matrix A is symmetric, and
16 Definitions and concepts

human PIN N = 5000 simple N = 5000 complex

1 1 1

%(λ)

0 0 0
-5 -3 -1 1 3 5 -5 -3 -1 1 3 5 -5 -3 -1 1 3 5

λ λ λ

Fig. 2.4 Adjacency matrix eigenvalue distributions %(λ|A) of the three graphs A shown and
quantified in previous figures. Left: Eigenvalue distribution for the human protein interaction
graph (see Figs. 1.1 and 2.3). Middle and right: Eigenvalue distributions for the two graphs
in Fig. 2.2. The middle histogram refers to the graph with only weak degree correlations
(top line in Fig. 2.2), and the right histogram refers to the graph with strong degree correla-
tions (bottom line in Fig. 2.2). The eigenvalue spectra of the last two graphs are seen to be
significantly different, in spite of their nearly identical degree distributions.

hence it will have N real-valued eigenvalues {λ1 (A), . . . , λN (A)} (which need not be
distinct). One defines the distribution of these eigenvalues as
N
1 X
%(λ|A) = δ[λ − λi (A)] (2.17)
N i=1

(see Appendix A for the definition and properties of the δ-distribution). It is easy
to prove that the smallest and largest eigenvalues λmin (A) and λmax (A) of sim-
ple nondirected graphs A obey various inequalities related to the degrees, such as
λmin (A) ≤ k̄(A) or k̄(A) ≤ λmax (A) ≤ maxi=1...N ki (A). The eigenvalue spectrum
of a nondirected graph contains information on the numbers of closed paths of all
possible lengths in this graph, since for any m ∈ IN we may write, using some simple
linear algebra,
Z ∞ N N
1 X m 1 X m
dλ λm %(λ|A) = λi (A) = (A )ii
−∞ N i=1 N i=1
N N N
1 X X X
= ... Ai i Ai i . . . Aim−1 im Aim i1 (2.18)
N i =1 i =1 i =1 1 2 2 3
1 2 m

Apart from the factor N , the last line gives us the total number of closed walks of
length m in the graph A (including self-intersecting and backtracking ones). Some
examples of adjacency matrix eigenvalue spectra for nondirected graphs are shown in
Fig. 2.4. For further examples, and more details on the mathematical relation between
 Solutions of exercises 17

the adjacency matrix spectra of nondirected random graphs and characteristics such
as their degree distributions and degree correlations, see, e.g. [158].

2.3 Solutions of exercises

Exercise 2.1
P
One calculates CN (which follows from the normalization requirement k≥0 p(k) = 1),
hki and hk 2 i using the geometric series `≥0 x` = (1−x)−1 :
P

X X
CN = e−γk = (e−γ )k = (1−e−γ )−1 (2.19)
k≥0 k≥0

1 d X −γ k 1 d d e−γ
hki = − (e ) = − CN = − log CN = (2.20)
CN dγ CN dγ dγ 1−e−γ
k≥0

1 d2 X −γ k 1 d2 1 d e−γ
hk 2 i = (e ) = CN = −
CN dγ 2 CN dγ 2 CN dγ (1−e−γ )2
k≥0
−γ
e (1−e−γ )2 + 2e−2γ (1−e−γ ) e−γ (1+e−γ )
= = (2.21)
(1−e−γ )3 (1−e−γ )2

Hence, the degree variance is

e−γ (1+e−γ ) e−2γ e−γ

σk2 = − = (2.22)
(1−e−γ )2 (1−e−γ )2 (1−e−γ )2

Next we turn to the scale-free degree

PN distribution. Again, CN follows from probabil-
ity normalization, giving CN = k=1 k −γ . In the limit N → ∞ the series k k −γ
P
converges if and only if γ > 1. The formulae for hki and the variance σk2 become

N
X N
X  N
X 2
−1 −1
hki = CN k 1−γ , σk2 = CN k 2−γ − CN k 1−γ (2.23)
k=1 k=1 k=1

The series for hki converges for N → ∞ if γ > 2. The series for hk 2 i (which is the first
term of σk2 ) converges if γ > 3. Hence the degree variance is finite if and only if γ > 3.
Now we choose γ = 2.5 and N = 104 , and approximate the summations as indicated:
N Z N
X
−λ 1 h 1−λ iN 1  1−λ 
k ≈ dk k−λ = k = N −1
1 1−λ 1 1−λ
k=1

With this we can calculate CN (by choosing λ = γ), hki (by choosing λ = γ − 1), and
hk 2 i (by choosing λ = γ − 2):
1 2−γ

1
2−γ N −1 N−2 − 1 1 − 0.01
hki ≈ 1
= 3 −3 =3 ≈ 2.97
1−γ
N 1−γ − 1
N 2 −1 1 − 0.000001
18 Definitions and concepts

1
N2 −1 99
hk 2 i = −3 3 =3 ≈ 297.00
N−2 − 1 1 − 0.000001

Hence the degree variance is approximately σk2 ≈ 297 − 9 = 288. The degree distri-
bution thus has average hki ≈ 3 and width σk ≈ 17. Although the width of power-law
distributed degree distributions with 2 < γ < 3 diverges as N → ∞, we see that for
realistic network sizes this divergence is a red herring. 2

Exercise 2.2
We start with the lower bound for (2.16). Since both Aij ∈ {0, 1} and δxi ,xj ∈ {0, 1},
we may write
N
1 X ki (A)kj (A)
Q(A) ≥ − δxi ,xj
N k̄(A) i,j=1 N k̄(A)
N
1 X ki (A)kj (A) 1 N 2 k̄ 2 (A)
≥− =− = −1 (2.24)
N k̄(A) i,j=1 N k̄(A) N k̄(A) N k̄(A)

The upper bound for (2.16) is proved in an even simpler way:

N N
1 X 1 X N k̄(A)
Q(A) ≤ Aij δxi ,xj ≤ Aij = =1 (2.25)
N k̄(A) i,j=1 N k̄(A) i,j=1 N k̄(A)

The upper bound is seen to be satisfied, for instance, by the trivial choice Aij = 1 for
all (i 6= j) and xi = xj for all (i, j). 2
3
Random graph ensembles

Definition and use of graph ensembles. A graph is defined in full by giving the number
N of its nodes and its adjacency matrix A; hence, in order to specify a random N -node
2
graph, we must give the set G ⊆∈ {0, 1}N of allowed graphs (the configuration space),
together with a probability distribution p(A) over this set1 . This combination {G, p}
of a graph set G with associated probabilities is called a random graph ensemble.
Apart from their use in formalizing what we mean by random graphs, thinking in
terms of random graph ensembles is very fruitful in the modelling of large complex
systems. In such systems, it is usually hard to measure and incorporate all microscopic
parameters (e.g. topology, forces) in full detail. Luckily, macroscopic characteristics can
often be written as sums over a huge number of microscopic variables, and here statisti-
cal laws tend to take over: if the system is sufficiently large2 , the values of macroscopic
characteristics can often be calculated with high accuracy already from the distribu-
tion of the microscopic parameters, since all microscopic states with non-negligible
probabilities give nearly identical values. This property is called self-averaging, and
is very common in statistical physics. For example, one of the key results of kinetic
theory is that the macroscopic properties of a gas, such as its pressure, do not depend
on the detailed motion of each individual particle, but only on the distribution of their
velocities, which, in equilibrium, is given by Maxwell’s formula. This formula only
depends on a few measurable control parameters, such as the temperature of the gas.
Similarly, in complex networks a detailed knowledge of all the microscopic entries
{Aij } of the adjacency matrix is often not accessible, and not even useful for the
purpose of efficiently extracting information about the macroscopic system behaviour.
Hence, one normally assumes a suitable probability distribution p(A) for the micro-
scopic variables A = {Aij }, and one calculates average values of macroscopic quantities
over this distribution. These averages over p(A) are expected to give predictions that
are correct even for individual graphs generated from p(A), if N is sufficiently large.
Like the Maxwell distribution, one expects moreover that suitable distributions p(A)
can be written in terms of only a few parameters – details should not matter!
The crucial question will now be how to choose suitable probabilities p(A), i.e.
how to ‘tailor’ random graph ensembles to mimic real-world graphs. In this section, we
define families of networks which satisfy specified topological conditions. This includes
a theoretical treatment of how to tailor an ensemble to meet hard or soft constraints.

2
1 Equivalently, we could always take G = {0, 1}N and assign p(A) = 0 to all disallowed graphs
A. Which route to choose is a matter of taste.
2 In practice this can even mean just a few hundreds of interacting variables.

Generating Random Networks and Graphs. First Edition. A.C.C. Coolen, A. Annibale, and
E.S. Roberts. c A.C.C. Coolen, A. Annibale, and E.S. Roberts 2017. Published in 2017 by
Oxford University Press. DOI: 10.1093/acprof:oso/9780198709893.001.0001
20 Random graph ensembles

3.1 The Erdös–Rényi graph ensemble

Overview: In this section, we define the simplest random graph ensemble, the Erdös–
Rényi model (ER). We derive some of its properties to gain intuition and as a warm-
up.

Definition. The classical and simplest random graph ensemble is the one proposed by
Erdös and Rényi [152]. Here G is the set of all simple nondirected N -node graphs, and
the 12 N (N −1) individual links Aij between the pairs of nodes (i, j) are for all A ∈ G
drawn independently, according to Prob(Aij = 1) = p and Prob(Aij = 0) = 1 − p:
Y 
p(A) = p δAij ,1 + (1 − p)δAij ,0 (3.1)
i<j

Since graphs generated according to this prescription will have on average, a mean
degree hki = p(N − 1), we can also write the graph probabilities (3.1) equivalently as
Y h hki  hki  i
p(A|hki) = δAij ,1 + 1 − δAij ,0 (3.2)
i<j
N −1 N −1

Note that hki is the average over the ensemble of the mean connectivity k̄(A) of the
individual graphs, i.e. hki = hk̄(A)i, where the average h. . .i is taken over (3.1).

Exercise 3.1 Show that the mean degree k̄(A) of individual N -node graphs a drawn from
the ER ensemble (3.1) is on average equal to p(N −1).

We see furthermore that, as a consequence of Aij ∈ {0, 1}, the probabilities (3.1)
can be written in the alternative form:
1
Y P P
pAij (1 − p)1−Aij = p i<j Aij (1 − p) 2 N (N−1)− i<j Aij

p(A) = (3.3)
i<j

From this, it follows that

Pthe dependence of p(A) on A can be expressed fully in terms
of the number L(A) = i<j Aij of links in the graph A, via the binomial formula

p(A) = pL(A) (1 − p)N(N −1)/2−L(A) (3.4)

Hence, we may conclude that the Erdös–Rényi ensemble assigns equal probabilities to
all graphs with the same number of links.

Exercise 3.2 Use thePself-averaging assumption to show that for N → ∞ the degree distri-
bution p(k|A) = N −1 i δki (A),k of an N -node graph drawn from the Erdös–Rényi ensemble
(3.2), in which hki remains fixed and finite3 , will become the Poissonian distribution

p(k) = e−hki hkik /k! (3.5)

3 This specific scaling regime of parameters, where hki is kept finite and N → ∞, is called the
sparse or finite connectivity regime.
 The Erdös–Rényi graph ensemble 21

0.9
hki = 2.0
0.8

0.7

0.6
F (d) hki = 1.8
0.5

0.4 hki = 1.6

0.3

0.2 hki = 1.4

0.1 hki = 1.2

0 ← hki = 1.0, 0.8, 0.6, 0.4, 0.2

0 20 40 60 80 100

Fig. 3.1 The fraction F (d) of node pairs (i, j) in randomly generated ER graphs that are
found to have a distance dij (A) ≤ d, shown versus d. Here N = 1000 and hki ∈ [0, 2]. One
observes clearly that there is a sharp transition at hki ≈ 1. For hki < 1, the fraction of node
pairs at a finite distance from each other is vanishingly small. For hki > 1, this fraction
suddenly becomes finite, which marks the emergence of the giant component in the graph.

Clustering and degree correlations. The average clustering coefficient Ci = hCi (A)i for
graphs from the ER ensemble (3.1) is found to be (see Appendix B):

Ci = p 1 − (1−p)N−1 − (N −1)p(1−p)N−2
 
(3.6)

If we keep the parameter p > 0 finite and fixed, while sending N → ∞, we obtain

lim Ci = p (3.7)
N →∞

If, alternatively, we set p = hki/(N −1) as in definition (3.2), and keep hki finite and
fixed while taking N → ∞, we find

hki h hki N −1 hki N −2 i

Ci = 1 − (1− ) − hki(1− )
N −1 N −1 N −1
hki h i
= 1 − e−hki − hkie−hki + O(N −2 ) (3.8)
N
So if hki is forced to be finite, then all clustering coefficients of typical Erdös–Rényi
graphs will vanish for N → ∞. Similarly, one can show (see again Appendix B) that
22 Random graph ensembles

large random graphs generated from the Erdös–Rényi ensemble (3.2) with fixed hki
will on average have vanishing degree correlations,
for all (k, k 0 ) : lim hW (k, k0 |A)i = hW (k|A)ihW (k 0 |A)i (3.9)
N →∞

with W (k, k 0 |A) as defined in (2.9).

Path lengths and the giant component. A further prominent feature of the the Erdös–
Rényi ensemble (3.2) with fixed hki is that for large N it exhibits a percolation tran-
sition at hki = 1. For values hki < 1, the typical graphs generated from the ensemble
will be highly fractured, and this is reflected in the observation that for the majority
of node pairs (i, j) there will not be any path in the graph that leads from i to j. In
contrast, as soon as hki > 1, a finite fraction of the N nodes will be connected, in what
is called a giant component. This effect can be observed, for instance, by measuring
the fraction F (d) of node pairs (i, j) in the graph that have as their shortest mutual
distance dij (A) ≤ d. Here dij (A) is the smallest value of ` ∈ IN (if any) such that
(A` )ij > 0; if i and j are not connected, we put dij (A) = ∞. See Figure 3.1, where
N = 1000. The percolation transition at hki = 1 becomes increasingly sharp for large
values of N .

Shortcut: The Erdös–Rényi ensemble is the simplest non-trivial graph ensemble, but
many of its properties are rather different from those observed in real (social, bio-
logical, commercial or technological) networks. The latter often have large clustering
coefficients, correlated degrees, so-called small-world properties, and non-Poissonian
scale-free degree distributions with fat tails.

3.2 Graph ensembles with hard or soft topological constraints

Overview: In this section, we present theoretical tools for defining random graph
ensembles systematically via soft or hard topological constraints.

Building topological features into graph ensembles. Given an observed network A? , we

now want to define systematically random graph ensembles {G, p} that are tailored
to generate graphs that are structurally similar to A? . To achieve this, we must first
define what we mean by similar. We quantify the relevant features that we want our
graphs to inherit from A? in terms of specific measurements Ωµ (A? ), µ = 1 . . . K,
and then demand either that these same measurements are reproduced by all graphs
generated in the ensemble (i.e. via a hard constraint), or that they are reproduced on
average (i.e. via a soft constraint) . Upon abbreviating the set of K chosen features as
Ω(A) = (Ω1 (A), . . . , ΩK (A)), this can be written as
hard constraint : all generated graphs must have the chosen features,
Ω(A) = Ω(A? ) for all A ∈ G with p(A) > 0 (3.10)

soft constraint : generated graphs must have the chosen features on average,
X
p(A)Ω(A) = Ω(A? ) (3.11)
A∈G
 Graph ensembles with hard or soft topological constraints 23

Our chosen features Ωµ (A) will normally represent topological characteristics of the
network that are important to the process that runs on it.
Both (3.10) and (3.11) give us conditions on the possible random graph ensembles,
but they do not yet specify the probabilities p(A) in full. The unique specification of
p(A) for both types of constraints will require information-theoretic arguments [53]. In
tailoring our random graphs we want to construct an ensemble {G, p} that incorporates
the chosen features Ω(A) but is otherwise strictly unbiased. Given a choice for the
set G, this will be true only for the distribution p(A) that, subject to the chosen
constraints (3.10) or (3.11), maximizes the Shannon entropy

X
S[p, G] = − p(A) log p(A) (3.12)
A∈G

The Shannon entropy defines the information content of a typical sample from the
distribution p(A). If we were to choose any distribution other than the one for which
the Shannon entropy is maximal (subject to the chosen constraints), we would by
definition already know something about our graphs in addition to the information
that is contained in the constraints. In other words: we would introduce an ad hoc bias
into the probability distribution p(A), and hence into our graph ensembles. We show
below how maximization of (3.12), subject to the applicable constraints, leads us to
unique formulas for our tailored graph ensembles.

Microcanonical ensembles: tailored random graphs with hard constraints. If we opt for
the hard constraint (3.10), we need only define the probabilities p(A) for all those
graphs A that obey Ω(A) = Ω(A? ). All other graphs are forbidden anyway. We
choose the collection of allowed graphs to be our set G, and we take the functions
Ωµ (A) to be integer-valued for simplicity (similar results can be obtained
P for non-
integer definitions). Maximization of (3.12) subject to the constraint A∈G p(A) = 1,
using Lagrange’s method, then gives the following equations that are to be solved
simultaneously for {p(A)} and the Lagrange parameter λ
∂ h  X i
S[p, G] − λ 1 − p(A0 ) = 0 (3.13)
∂p(A) 0 A ∈G

Working out the above partial derivatives leads to log p(A) = λ − 1 for all A ∈
G.
P The value of the Lagrange parameter λ follows from the normalization condition
A∈G p(A) = 1, and the end result is the following maximum entropy ensemble, in
which all graphs A that meet our hard constraints are equally probable
1
∀A ∈ G : p(A|Ω) = (3.14)
Z(Ω)
2
Z(Ω) ≥ 1 is the number of graphs in {0, 1}N that have Ω(A) = Ω(A? ). We could
2
also choose to work with the set G = {0, 1}N of all N -node graphs, and forbid those
graphs A that do not satisfy the constraints by assigning to them p(A) = 0. This gives
formulae of the form
24 Random graph ensembles

δΩ(A),Ω X
∀A ∈ G : p(A|Ω) = Z(Ω) = δΩ(A),Ω (3.15)
Z(Ω)
A∈G

Global constraints such as non-directionality or absence of self-interactions can in

principle be incorporated via the quantities Ωµ (A), but in practice it is often more
transparent to build such constraints into the definition of G. For instance, we can
limit ourselves to simple nondirected graphs by choosing in definition (3.15) the set
2
G = {A ∈ {0, 1}N | Aij = Aji for all (i, j) and Aii = 0 for all i} (3.16)
The hard-constrained ensembles (3.15) are also called microcanonical ensembles.

Canonical ensembles: tailored random graphs with soft constraints. If we opt to tailor
our ensemble via the soft topological constraints (3.11), we have to build P any global
constraints into the set G, and maximize the entropy (3.12) subject to A∈G p(A) = 1
?
P
and A∈G p(A)Ω µ (A) = Ωµ (A ) for all µ = 1 . . . K. The Lagrange maximization
method will now involve K + 1 Lagrange parameters {λµ }, K of which correspond
to
P the imposed constraints, and λ0 corresponds to the normalization requirement
A∈G p(A) = 1. Our equations for {p(A)} then become

K
∂ h  X  X  X i
S[p, G] − λ0 1 − p(A0 ) − λµ Ωµ (A? ) − p(A0 )Ωµ (A0 ) = 0
∂p(A) 0 0
A ∈G µ=1 A ∈G
(3.17)
The solution of these latter equations takes the form
PK
λµ Ωµ (A)
e µ=1 X PK
λµ Ωµ (A)
p(A) = , Z(Ω) = e µ=1 (3.18)
Z(Ω)
A∈G

in which the parameters {λ1 , . . . , λK } are found by solving the coupled equations
1 X PK
∀µ ∈ {1, . . . , K} : Ωµ (A? ) = e µ=1 λµ Ωµ (A) Ωµ (A). (3.19)
Z(Ω)
A∈G

One can show that equations (3.19) have a unique solution, so again we arrive at a
well-defined tailored random graph ensemble. Solving (3.19), whether analytically or
numerically, can unfortunately be quite difficult. It is sometimes possible to proceed
analytically, if one chooses the observables Ωµ (A) wisely and if N is large. The soft-
constrained ensembles (3.18) are also called exponential or canonical ensembles.
In the above derivations of the hard-constrained and the soft-constrained ensem-
bles, we should in principle have also included the inequality constraints p(A) ≥ 0
for all A ∈ G when maximizing the Shannon entropy (3.12) – after all, probabilities
are not allowed to be negative. However, it turned out in both cases that even with-
out imposing them explicitly, the inequality constraints are satisfied automatically by
the maximum entropy distributions we obtained. Hence, these further constraints are
obsolete.
 The link between ensembles and algorithms 25

Exercise 3.3 Show that the distribution p(A) which maximizes the ShannonP entropy (3.12)
for simple nondirected graphs, subject
P to the average degree constraint A∈G p(A)k̄(A) =
hki and subject to the normalization A∈G p(A) = 1, is the Erdös–Rényi ensemble (3.2).

Chapters 4 and 6 will explore soft- and hard-constrained ensembles in greater

detail, including the mechanism for including ‘hidden’ constraints and a discussion of
how to combine both hard and soft constraints simultaneously in a single ensemble.

Shortcut: To find the correct probability weights p(A) to define an ensemble of graphs
that satisfy certain (hard or soft) topological constraints
P but are otherwise unbiased, it
is necessary to maximize the Shannon entropy S = − A∈G p(A) log p(A). One finds
upon doing so that for hard constraints every valid configuration is given the same
probability weight 1/|G|, where |G| represents the number of allowed configurations. For
soft-constrained ensembles, the solution is more subtle, and may need to be approached
via numerical simulations or N → ∞ type limiting arguments.

3.3 The link between ensembles and algorithms

Overview: In this section, we discuss the link between graph ensembles, i.e. the statis-
tical specifications of desired topological features of our random graphs, and algorithms,
i.e. the dynamical processes for the evolution of graphs.

Markovian graph generation algorithms. A graph generation algorithm is a discrete

time4 dynamical process for the adjacency matrix A. Virtually all graph generation
algorithms are stochastic. The evolution involves at each step one or more random
numbers, so that one cannot write deterministic dynamical equations for the graph A
itself, but only equations relating to the probability of finding a certain graph at a given
time step. Most algorithms are of the so-called Markovian form, or can be transformed
into that form, which means that the graph probabilities at the next iteration step
depend only on those of the previous step (as opposed to depending also on states
at even earlier stages of the process). Deterministic algorithms are just special cases
of stochastic ones, in which the relevant state probabilities at any time can take only
the values 0 or 1. Some algorithms take the form of growth processes, where nodes
2
are being added dynamically, so not only the adjacency matrix A ∈ {0, 1}N but also
the graph size N itself evolves in time. It follows that the generic graph generation
problem can be summarized as in Fig. 3.2.
The quantities W (N 0, A0 ) → (N, A) ∈ [0, 1] in Fig. 3.2 represent the probabilities
 

to move to a new state (N, A) at the next iteration if the present state is (N 0, A0 ),
and they therefore contain the complete specification of an algorithm. Any applicable
hard topological constraints must be built into the algorithm by forbidding all moves
into states that
P violate
P these constraints. We note that probability conservation re-
quires that N≥0 A∈{0,1}N 2 W (N 0, A0 ) → (N, A) = 1 for all (N 0, A0 ). In addition


4 Time is inevitably discrete if an algorithm runs on a conventional digital computer.

26 Random graph ensembles

to choosing the transition probabilities, one also needs to initialize the algorithm, i.e.
one must specify p0 (N, A) for all (N, A), subject to the condition that p0 (N, A) = 0
2
unless A ∈ {0, 1}N . The stationary graph sampling weights p(N, A) to which an
algorithm evolves, together with the set G of graphs for which these probabilities will
be non-zero, will then define our random graph ensemble in the sense of the previous
section.
Note that the link between graph generation algorithms and the graph sampling
weights found upon their equilibration (the stationary probabilities) in Fig. 3.2 need
not be unique. For each choice of stationary probabilities there exist multiple stochas-
tic processes that would have these probabilities as stationary states; this is often a
welcome feature, since in practical applications where we just want to generate graphs
with specified probabilities, any process that equilibrates to our desired graph ensem-
ble would do. The more algorithms there are that will do our job, the easier it should
be to find and implement one. In contrast, if we start by defining a graph generation
algorithm, then any potential non-uniqueness of the associated stationary probabilities
is problematic. Here we have to ensure that our algorithm is ergodic, i.e. that any state
(N, A) that does not violate applicable hard constraints can in principle be reached by
our stochastic process with non-zero probability in a finite number of iteration steps.
Proving ergodicity of stochastic processes can be hard, but if ergodicity is violated
we run the danger that the graph probabilities after equilibration will depend on our
choice of initialization (which introduces a dangerous bias into our sampling).
The above picture represents a mathematical ideal that in practice cannot be re-
alized fully on a computer. First, no truly random numbers ever emerge from com-
puter code, only numbers that are approximately random. No numerical algorithm
will therefore be really random, and when graphs are large, one has to be careful that
the number of random number calls made by our code does not approach the period
of our random number generator. Secondly, true equilibration of a stochastic process
generally requires t → ∞, but in practice we terminate our algorithm after a specified
number of steps which we expect to be sufficient. Moreover, algorithms in which the
number of nodes increases have to be stopped by hand simply to end with a graph
of finite size. The choice of termination time has to be substantiated somehow, either
via analytical calculation of process relaxation times (which is usually hard), or via
monitoring of the Hamming distance5 between the graph A0 at time t = 0 and the
current graph At (for algorithms that keep the size N fixed). Only in very simple cases
will we work with graph generation algorithms that equilibrate truly within a finite
number of steps; for instance, one could interpret (3.1) trivially as a process that keeps
N fixed and converges within one iteration step.

Randomization versus growth. In writing this book, we have found it helpful to distin-
guish between the following two classes of graph generation algorithms, as they tend
to be used and studied in distinct communities and give rise to distinct questions:
• Graph randomization algorithms.
These are algorithms with W (N 0, A0 ) → (N, A) = 0 unless N = N 0 , so the
 

5 This is the number of entries for which the two matrices A0 and At differ.
 The link between ensembles and algorithms 27

GRAPH GENERATION ALGORITHM :

evolving probabilities
X X
pt+1 (N, A) = W [(N 0, A0 ) → (N, A)] pt (N 0, A0 )
N 0 ≥0 A0 ∈{0,1} N 02

GRAPH SAMPLING WEIGHTS :

stationary probabilities

p(N, A) = lim pt (N, A)

t→∞

Fig. 3.2 The central representation of graph generation as the result of a discrete time
Markovian stochastic process, i.e. a Markov chain. The core of the graph generation problem
is to establish the precise link between specified generation algorithms and the stationary
probabilities to which these algorithms will converge. This representation captures virtually
all algorithms in the literature today. It depends on the context of one’s study and the
preferences in one’s research community whether one enters this box from the top or from
the bottom. If we enter at the top, we define our random graphs by the algorithm, and then
need to know the associated stationary probabilities in order to anticipate the features of
the graphs that will be generated upon termination of our algorithm. If we enter from the
bottom, we specify our random graphs by the desired sampling weights (i.e. by the desired
random graph ensemble), and our problem is to construct an algorithm that will have these
weights as its stationary probabilities upon convergence.

number of nodes in the graph is never changed. One usually starts from some seed
2
graph A? ∈ {0, 1}N , and hard topological constraints are built in by choosing
a seed graph in which these constraints are satisfied, in combination with a set
of allowed process moves that leave invariant the values of the chosen topological
features to be constrained. Since N is fixed, we can now simplify the algorithm’s
2
Markov chain equation, using the implicit convention that always A, A0 ∈ {0, 1}N ,
to
X
pt+1 (A) = W[A0 → A] pt (A0 ), p0 (A) = δA,A? (3.20)
A0

• Graph growth algorithms.

These are algorithms with W (N 0, A0 ) → (N, A) = 0 unless N > N 0 . Here, at each
 

iteration step one increases the number of nodes in the graph, and one generates
new links to connect the new nodes to the existing ones. The initialization would
typically be a set of disconnected nodes.
28 Random graph ensembles

3.4 Solutions of exercises

Exercise 3.1
We have:
X 1 XX
hk̄(A)i = p(A)k̄(A) = p(A)ki (A)
N i
A∈G A∈G
1 XX
= p(A)Aij (3.21)
N
i6=j A∈G

Upon defining the single link probabilities

p(Ak` ) = pδAk` ,1 + (1 − p)δAk` ,0 (3.22)

we may write
1 X X hY i
hk̄(A)i = p(Ak` ) Aij
N
i6=j A∈G k<`
1 Xh X i Y h X i 1 X
= p(Aij )Aij × p(Ak` |p) = p
N N
i6=j Aij ∈{0,1} (k,`)6=(i,j) Ak` ∈{0,1} i6=j

= p(N − 1) (3.23)
P
Note
Q thatP in the penultimate line we used factorization over links of the sum A∈G =
, we isolated the contribution from Aij , and we used the two identities
Pk<` Ak` ∈{0,1}
2
P
Ak` ∈{0,1} p(A k` ) = 1 and Ak` ∈{0,1} p(Ak` )Ak` = p.

Exercise 3.2
The assumed self-averaging property of macroscopic observables such as p(k|A) for
large graphs allows us to write
X
lim p(k|A) = lim p(k|A)p(A|hki) (3.24)
N→∞ N→∞
A∈G

Using the integral representation of the Kronecker delta symbol (2.3), we obtain
1 X π dω iωk X
Z P
lim p(k|A) = lim e p(A|hki)e−iω ` Ai` (3.25)
N →∞ N →∞ N −π 2π
i a∈G

At
P this point
P we can useP the symmetry of the adjacency matrix and the identity
A
` i` = A δ
k` k` ki = k<` Ak` (δki + δ`i ) to write

1 X π dω iωk X
Z
1
P
lim p(k|A) = lim e p(A|hki)e− 2 iω k<` Ak` (δki +δ`i )(3.26)
N→∞ N→∞ N −π 2π
i A∈G

The next step is to insert the probabilities (3.2):

 Solutions of exercises 29

lim p(k|A)
N →∞
1 X π dω iωk X Y h hki hki

Z i
= lim e δAk`,1 + 1− δAk` ,0 e−iωAk` (δki +δ`i )
N→∞ N
i −π 2π N −1 N −1
A∈G k<`
π
hki −iω(δki +δ`i ) hki

Z
1 X dω iωk Y X h i
= lim e e δAk`,1 + 1− δAk` ,0
N→∞ N
i −π 2π N −1 N −1
k<` Ak` ∈{0,1}

1 X π dω iωk Y h hki
Z

i
= lim e 1+ e−iω(δki +δ`i ) − 1 (3.27)
N→∞ N
i −π 2π N −1
k<`

where the last two steps were achieved by multiplying through the exponential, and
eliminating terms that must certainly be equal to 1 by the action of the Kronecker
delta function. The next stage uses the expansion 1 + x = exp(x − 12 x2 + O(x3 )) for
x'0

lim p(k|A)
N →∞
1 X π dω iωk Y Nhki
Z
−iω(δki +δ`i ) hki2
−1)− (e−iω(δki +δ`i ) −1)+O(N −3 )
= lim e e −1 (e 2N 2
N→∞ N −π 2π
i k<`
1 X π dω iωk Y (δki +δ`i ) Nhki
Z
hki2
 
−iω
−1 (e −1)− 2N 2 (e−iω −1) +O(N −3 )
= lim e e
N→∞ N −π 2π
i k<`
Z π
1 dω iωk+ 12 P`6=k (δki +δ`i ) Nhki hki2
 
−iω
−1)− 2N 2 (e−iω −1) +O(N −3 )
X
= lim e −1 (e
N→∞ N −π 2π
i
Z π
1 X dω iωk+hki(e−iω −1)+O(N −1 )
= lim e (3.28)
N→∞ N −π 2π
i

After writing the factor exp[hkie−iω ] as a Taylor series, we then obtain a form in which
we can recognize and exploit once more the integral representation of the Kronecker
delta symbol:
Z π
dω iωk X hkir −irω X hkir
lim p(k|A) = e−hki e e = e−hki δk,r
N→∞ −π 2π r≥0
r! r!
r≥0
−hki k
=e hki /k! (3.29)

which gives the Poissonian degree distribution (3.5), as claimed. 2

30 Random graph ensembles

Exercise 3.3
We apply the general results (3.18, 3.19)
P for soft-constrained maximum entropy ensem-
bles. Upon selecting k̄(A) = N −1 i ki (A) for our observable Ω(A), we find (3.18)
reducing to
−1 PN
eλN i,j=1 Aij X −1 PN
p(A) = , Z(hki) = eλN i,j=1 Aij
(3.30)
Z(hki)
A∈G

We can rewrite this probability distribution, using the symmetry of A:

−1 P
e2λN i<j Aij 1 Y h 2λ/N i
p(A) = = e δAij ,1 + δAij ,0 (3.31)
Z(hki) Z(hki) i<j

The normalization sum Z(hki) becomes

X Yh i
Z(hki) = e2λ/N δAij ,1 + δAij ,0
A∈G i<j
Y X h i Yh i
= e2λ/N δAij ,1 + δAij ,0 = 1 + e2λ/N (3.32)
i<j Aij ∈{0,1} i<j

In combination we can now write

Y h e2λ/N 1 i
p(A) = δA ,1 + δA ,0 (3.33)
i<j
1 + e2λ/N ij 1 + e2λ/N ij

The equation (3.19) from which to solve λ simplifies to

1 XX Y h e2λ/N 1 i
hki = A`r 2λ/N
δAij ,1 + 2λ/N
δAij ,0
N 1+e 1+e
A∈G
`r i<j

2 X X h e2λ/N 1 i
= A`r δ A `r ,1 + δ A `r ,0
N 1 + e2λ/N 1 + e2λ/N
`<r A`r ∈{0,1}

2 X e2λ/N e2λ/N
= = (N −1) (3.34)
N 1 + e2λ/N
`<r
1 + e2λ/N

This allows us to eliminate λ and show that (3.33) is indeed equivalent to

Y h hki hki
i
p(A) = δAij ,1 + 1 − δAij ,0 (3.35)
i<j
N −1 N −1

2
Part II
Random graph ensembles
4
Soft constraints: exponential random
graph models

Introduction. Exponential random graphs are those sampled from ensembles of the
canonical form (3.18), which was derived by combining maximum entropy arguments
with soft topological constraints, and involves graph probabilities p(A) that depend
exponentially on the quantities used to formulate the constraints.
Exponential random graph models (ERGMs) provide conceptually simple recipes
for generating random graphs and for estimating the values of observables of random
graphs, given only the values of some other observables. However, they are not without
difficulties. While they work well for linear constraints, such as an imposed degree
sequence, they can exhibit degenerate behaviour in more complicated cases. In this
chapter, we illustrate such behaviour, focusing on the phenomenology of so-called phase
transitions. These are abrupt quantitative changes in the macroscopic characteristics
of the sampled graphs, which become increasingly prominent as the graph size N
increases, and which make it very difficult in practice to control the values of relevant
ensemble averages in the required manner.
We work out several examples of ERGMs, hopefully in enough detail to not only
provide an understanding of the models and their limitations, but also to give the
reader the opportunity to extend these results to develop their own bespoke ERGMs.
We also show that the ERGM approach can be used to define network ensembles with
community structure, the so-called stochastic block models (SBMs).

4.1 Definitions and basic properties of ERGMs

Overview: We define ERGMs as soft-constrained ensembles with parametrized prob-
abilities p(A), in which we exploit the exponential dependence on the parameters.

ERGMs as exponentially parametrized ensembles. In the previous chapter, we intro-

duced the concept of network or graph ensembles. Each ensemble is defined by a set G
of all allowed microscopic graph realizations A, together with a probability distribution
p(A) over the set. Exponential random graph models are those where the probabili-
ties are of the general canonical form (3.18). They were first introduced in [88], and
further developed in [76] and in several later studies such as [170, 181, 159, 13, 165].
They are nowadays widely used, especially in the social sciences. Computational tools
to analyze and simulate networks based on ERG models are widely available on the
Internet.
Generating Random Networks and Graphs. First Edition. A.C.C. Coolen, A. Annibale, and
E.S. Roberts. c A.C.C. Coolen, A. Annibale, and E.S. Roberts 2017. Published in 2017 by
Oxford University Press. DOI: 10.1093/acprof:oso/9780198709893.001.0001
34 Soft constraints: exponential random graph models

In ERGMs we often regard the probabilities p(A) as parametrized directly by

λ = {λ1 , . . . , λK }. We no longer view these latter parameters as unknown complicated
functions of the values Ω of the soft-constrained topological features Ω(A) that have
to be solved and eliminated via (3.11), and we simply write
PK
λµ Ωµ (A)
e µ=1 X PK
λµ Ωµ (A)
p(A) = , Z(λ) = e µ=1 (4.1)
Z(λ)
A∈G

The graph set G will normally only code for basic topological features, such as non-
directedness. We have seen earlier that ensembles of the form P (4.1) are maximally
random, in the sense of maximizing the Shannon entropy S = − A∈G p(A) log p(A)
given the imposed values of the ensemble averages of {Ω1 (A), . . . , ΩK (A)}. In this
chapter, we focus on those properties of ERGMs that derive from their exponential
form alone, and various subtleties which one will encounter when sampling from such
ensembles.

Averages, fluctuations and susceptibilities. Once the parameter values λ of an ERGM

are known, one can use the corresponding probability distribution p(A) to estimate
the value of any observable f (A) by calculating the ensemble average
PK
f (A)e µ=1 λµ Ωµ (A)
P
A∈G
hf (A)i = (4.2)
Z(λ)

In particular, the expectation values for the key observables Ωµ (A) can all be written
as partial derivatives of the quantity F (λ) = log Z(λ), which apparently acts as a
generating function1 :
1 X P
hΩµ (A)i = Ωµ (A)e ν λν Ων (A)
Z(λ)
A∈G
1 ∂ ∂
= Z(λ) = F (λ) (4.3)
Z(λ) ∂λµ ∂λµ

The same turns out to be true for the fluctuations in the key observables, since
1 X 2 P  1 ∂ 2
hΩ2µ (A)i − hΩµ (A)i2 = Ωµ (A)e ν λν Ων (A) − Z(λ)
Z(λ) Z(λ) ∂λµ
A∈G
2
1 ∂ 2 Z(λ)
 
1 ∂Z(λ)
= −
Z(λ) ∂λ2µ Z 2 (λ) ∂λµ
2
 
∂ 1 ∂Z(λ) ∂ F (λ)
= = (4.4)
∂λµ Z(λ) ∂λµ ∂λ2µ

1 In the terminology of statistical physics the quantity F (λ) would, apart from an overall mul-
tiplicative constant, be called the free energy. Similarly, one would call the normalizing constant
Z(λ) the partition function and the exponent of the probabilities, apart from an overall factor, the
Hamiltonian of the system: H(A) = − K
P
µ=1 Ωµ (A).
Other documents randomly have
different content
shapes of statues, the radiance of carefully selected vestments with
hues inwrought upon pure marble. The rhythms and melodies of the
Doric mood were sympathetic to the proportions of the Doric
colonnades. The grove of pillars through which the pageant passed
grew from the living rock into shapes of beauty, fulfilling by the
inbreathed spirit of man Nature’s blind yearning after absolute
completion. The sun himself—not thwarted by artificial gloom, or
tricked with alien colours of stained glass—was made to minister in
all his strength to a pomp, the pride of which was a display of form in
manifold magnificence. The ritual of the Greeks was the ritual of a
race at one with Nature, glorying in its affiliation to the mighty mother
of all life, and striving to add by human art the coping-stone and final
touch to her achievement.

Sketches in Italy and Greece (London, 1874).

 THE CATHEDRAL OF ROUEN.
THOMAS FROGNALL DIBDIN.

THE approach to Rouen is indeed magnificent. I speak of the

immediate approach; after you reach the top of a considerable rise,
and are stopped by the barriers, you then look down a straight,
broad, and strongly paved road, lined with a double row of trees on
each side. As the foliage was not thickly set, we could discern,
through the delicately clothed branches the tapering spire of the
Cathedral and the more picturesque tower of the Abbaye St. Ouen—
with hanging gardens and white houses to the left—covering a richly
cultivated ridge of hills, which sink as it were into the Boulevards and
which is called the Faubourg Cauchoise. To the right, through the
trees, you see the river Seine (here of no despicable depth or
breadth) covered with boats and vessels in motion: the voice of
commerce and the stir of industry cheering and animating you as
you approach the town. I was told that almost every vessel which I
saw (some of them two hundred and even of three hundred tons
burthen) was filled with brandy and wine. The lamps are suspended
from the centre of long ropes, across the road; and the whole scene
is of a truly novel and imposing character. But how shall I convey to
you an idea of what I experienced, as, turning to the left, and leaving
the broader streets which flank the quay, I began to enter the
penetralia of this truly antiquated town? What narrow streets, what
overhanging houses, what bizarre, capricious ornaments! What a
mixture of modern with ancient art! What fragments, or rather what
ruins of old delicately-built Gothic churches! What signs of former
and of modern devastation! What fountains, gutters, groups of never-
ceasing men, women and children, all occupied, and all apparently
happy! The Rue de la Grosse Horloge (so called from a huge,
clumsy, antiquated clock which goes across it) struck me as being
not among the least singular streets of Rouen. In five minutes I was
within the courtyard of the Hôtel Vatel, the favourite residence of the
English.
It was evening when I arrived in company with three Englishmen.
We were soon saluted by the laquais de place—the leech-like
hangers-on of every hôtel—who begged to know if we would walk
upon the Boulevards. We consented; turned to the right; and,
gradually rising gained a considerable eminence. Again we turned to
the right, walking upon a raised promenade; while the blossoms of
the pear and apple trees, within a hundred walled gardens, perfumed
the air with a delicious fragrance. As we continued our route along
the Boulevard Beauvoisine, we gained one of the most interesting
and commanding views imaginable of the city of Rouen—just at that
moment lighted up by the golden rays of a glorious sunset—which
gave a breadth and a mellower tone to the shadows upon the
Cathedral and the Abbey of Saint Ouen....
CATHEDRAL OF ROUEN
 I have now made myself pretty well acquainted with the
geography of Rouen. How shall I convey to you a summary, and yet
a satisfactory description of it? It cannot be done. You love old
churches, old books, and relics of ancient art. These be my themes,
therefore: so fancy yourself either strolling leisurely with me, arm in
arm, in the streets—or sitting at my elbow. First for the Cathedral:—
for what traveller of taste does not doff his bonnet to the Mother
Church of the town through which he happens to be travelling—or in
which he takes up a temporary abode? The west front, always the
forte of the architect’s skill, strikes you as you go down, or come up,
the principal street—La Rue des Calmes,—which seems to bisect
the town into two equal parts. A small open space (which, however
has been miserably encroached upon by petty shops) called the
Flower-garden, is before this western front; so that it has some little
breathing room in which to expand its beauties to the wondering
eyes of the beholder. In my poor judgment, this western front has
very few elevations comparable with it—including even those of
Lincoln and York. The ornaments, especially upon three porches,
between the two towers, are numerous, rich, and for the greater part
17
entire:—in spite of the Calvinists, the French Revolution, and time.
Among the lower and smaller basso-relievos upon these porches is
the subject of the daughter of Herodias dancing before Herod. She is
manœuvering on her hands, her feet being upwards. To the right, the
decapitation of Saint John is taking place.
The southern transept makes amends for the defects of the
northern. The space before it is devoted to a sort of vegetable
market: curious old houses encircle this space: and the ascent to the
door, but more especially the curiously sculptured porch itself, with
the open spaces in the upper part—light, fanciful and striking to a
degree—produce an effect as pleasing as it is extraordinary. Add to
this the ever-restless feet of devotees, going in and coming out, the
worn pavement, and the frittered ornaments, in consequence—seem
to convince you that the ardour and activity of devotion is almost
equal to that of business.
As you enter the Cathedral, at the centre door, by descending
two steps, you are struck with the length and loftiness of the nave,
and with the lightness of the gallery which runs along the upper part
of it. Perhaps the nave is too narrow for its length. The lantern of the
central large tower is beautifully light and striking. It is supported by
18
four massive clustered pillars, about forty feet in circumference;
but on casting your eye downwards, you are shocked at the
tasteless division of the choir from the nave by what is called a
Grecian screen: and the interior of the transepts has undergone a
like preposterous restoration. The rose windows of the transepts,
and that at the west end of the nave, merit your attention and
commendation. I could not avoid noticing to the right, upon entrance,
perhaps the oldest side chapel in the Cathedral: of a date, little less
ancient than that of the northern tower, and perhaps of the end of the
Twelfth Century. It contains by much the finest specimens of stained
glass—of the early part of the Sixteenth Century. There is also some
beautiful stained glass on each side of the Chapel of the Virgin,
behind the choir; but although very ancient, it is the less interesting,
as not being composed of groups, or of historical subjects. Yet, in
this, as in almost all the churches which I have seen, frightful
devastations have been made among the stained-glass windows by
the fury of the Revolutionists....
As you approach the Chapel of the Virgin, you pass by an
ancient monument, to the left, of a recumbent Bishop, reposing
behind a thin pillar, within a pretty ornamented Gothic arch. To the
eye of a tasteful antiquary this cannot fail to have its due attraction.
While, however, we are treading upon hallowed ground, rendered if
possible more sacred by the ashes of the illustrious dead, let us
move gently onwards towards the Chapel of the Virgin, behind the
choir. See, what bold and brilliant monumental figures are yonder to
the right of the altar! How gracefully they kneel and how devoutly
they pray! They are the figures of the Cardinals D’Amboise—uncle
and nephew:—the former minister of Louis XII. and (what does not
necessarily follow, but what gives him as high a claim upon the
gratitude of posterity) the restorer and beautifier of the glorious
building in which you are contemplating his figure. This splendid
monument is entirely of black and white marble, of the early part of
the Sixteenth Century. The figures just mentioned are of white
marble, kneeling upon cushions, beneath a rich canopy of Gothic
fret-work....
The south-west tower remains, and the upper part of the central
tower, with the whole of the lofty wooden spire:—the fruits of the
liberality of the excellent men of whom such honourable mention has
been made. Considering that this spire is very lofty, and composed
of wood, it is surprising that it has not been destroyed by tempest or
19
by lightning. The taste of it is rather capricious than beautiful....
Leaving the Cathedral, you pass a beautifully sculptured fountain
(of the early time of Francis I.) which stands at the corner of a street,
to the right; and which, from its central situation, is visited the live-
long day for the sake of its limpid waters.

A Bibliographical, Antiquarian, and Picturesque Tour in France

and Germany (London, 1829).
 THE CASTLE OF HEIDELBERG.
VICTOR HUGO.

THERE is every style in the Castle of Heidelberg. It is one of those

buildings where are accumulated and mingled beauties which
elsewhere are scattered. It has some notched towers like
Pierrefonds, some jewelled façades like Anet, some fosse-walls
fallen into the moat in a single piece like Rheinfels, some large
sorrowful fountains, moss-grown and ready to fall, like the Villa
Pamfili, some regal chimney-pieces filled with briers and brambles,—
the grandeur of Tancarville, the grace of Chambord, the terror of
Chillon....
 THE CASTLE OF HEIDELBERG.

If you turn towards the Palace of Frederick IV. you have before
you the two high, triangular pediments of this dark and bristling
façade, the greatly projecting entablatures, where, between four
rows of windows, are sculptured with the most spirited chisel, nine
Palatines, two Kings, and five Emperors.
On the right you have the beautiful Italian front of Otho-Heinrich
with its divinities, its chimeræ, and its nymphs who live and breathe
velveted by the soft shadows, with its Roman Cæsars, its Grecian
demi-gods, its Hebraic heroes, and its porch which was sculptured
by Ariosto. On the left you catch a glimpse of the Gothic front of
Louis the Bearded, as savagely dug out and creviced as if gored by
the horns of a gigantic bull. Behind you, under the arches of a porch,
which shelters a half-filled well, you see four columns of grey granite,
presented by the Pope to the great Emperor of Aix-la-Chapelle,
which in the Eighth Century went to Ravenna on the border of the
Rhine, in the Fifteenth, from the borders of the Rhine to the borders
of the Neckar, and which, after having witnessed the fall of
Charlemagne’s Palace at Ingelheim, have watched the crumbling of
the Palatines’ Castle at Heidelberg. All the pavement of the court is
covered with ruins of flights of steps, dried-up fountains, and broken
basins. Everywhere the stones are cracked and nettles have broken
through.
The two façades of the Renaissance which give such an air of
splendour to this court are of red sandstone and the statues which
decorate them are of white sandstone, an admirable combination
which proves that the great sculptors were also great colourists.
Time has rusted the red sandstone and given a golden tinge to the
white. Of these two façades one, that of Frederick IV., is very severe;
the other, that of Otho Heinrich, is entirely charming. The first is
historical, the second is fabulous. Charlemagne dominates the one,
Jupiter dominates the other.
The more you regard these two Palaces in juxtaposition and the
more you study their marvellous details, the more sadness gains
upon you. Strange destiny for masterpieces of marble and stone! An
ignorant visitor mutilates them, an absurd cannon-ball annihilates
them, and they were not mere artists but kings who made them.
Nobody knows to-day the names of those divine men who built and
sculptured the walls of Heidelberg. There is renown there for ten
great artists who hover nameless above this illustrious ruin. An
unknown Boccador planned this Palace of Frederick IV.; an ignored
Primaticcio composed the façade of Otho-Heinrich; a Cæsar
Cæsarino, lost in the shadows, designed the pure arches to the
equilateral triangle of Louis V.’s mansion. Here are arabesques of
Raphael, and here are figurines of Benvenuto. Darkness shrouds
everything. Soon these marble poems will perish,—their poets have
already died.
For what did these wonderful men work? Alas! for the sighing
wind, for the thrusting grass, for the ivy which has come to compare
its foliage with theirs, for the transient swallow, for the falling rain,
and for the enshrouding night.
 One singular thing here is that the three or four bombardments
to which these two façades have been subjected have not treated
them in the same way. Only the cornice and the architraves of Otho-
Heinrich’s Palace have been damaged. The immortal Olympians
who dwell there have not suffered. Neither Hercules, nor Minerva,
nor Hebe has been touched. The cannon-balls and shells crossed
each other here without harming these invulnerable statues. On the
other hand, the sixteen crowned knights, who have heads of lions on
the grenouillières of their armour and who have such valiant
countenances, on the Palace of Frederick IV. have been treated by
the bombs as if they had been living warriors. Nearly every one of
them has been wounded. The face of the Emperor Otho has been
covered with scars; Otho, King of Hungary, has had his left leg
fractured; Otho-Heinrich, the Palatine, has lost his hand; a ball has
disfigured Frederick the Pious; an explosion has cut Frederick II. in
half and broken Jean Casimir’s loins. In the assaults which were
levelled at the highest row, Charlemagne has lost his globe and in
the lower one Frederick IV. has lost his sceptre.
However, nothing could be more superb than this legion of
princes all mutilated and all standing. The anger of Leopold II. and of
Louis XIV., the thunder—the anger of the sky, and the anger of the
French Revolution—the anger of the people, have vainly assailed
them; they all stand there defending their façade with their fists on
their hips, with their legs outstretched, with firmly planted heel and
defiant head. The Lion of Bavaria is proudly scowling under their
feet. On the second row beneath a green bough, which has pierced
through architrave and which is gracefully playing with the stone
feathers of his casque, Frederick the Victorious is half drawing his
sword. The sculptor has put into his face an indescribable
expression of Ajax challenging Jupiter and Nimrod shooting his
arrow at Jehovah. These two Palaces of Otho-Heinrich and
Frederick IV. must have offered a superb sight when seen in the light
of that bombardment on the fatal night of May 21, 1693....
To-day the Tower of Frederick the Victorious is called the Blown-
up Tower.
 Half of this colossal cylinder of masonry lies in the moat. Other
cracked blocks detached from the top of the tower would have fallen
long ago if the monster-trees had not seized them in their powerful
claws and held them suspended above the abyss.
A few steps from this terrible ruin chance has made a ruin of
ravishing beauty; this is the interior of Otho-Heinrich’s Palace, of
which until now I have only described the façade. There it stands
open to everybody under the sunshine and the rain, the snow and
the wind, without a ceiling, without a canopy, and without a roof,
whose dismantled walls are pierced as if by hazard with twelve
Renaissance doors,—twelve jewels of orfèvrerie, twelve chefs
d’œuvre, twelve idyls in stone—entwined as if they issued from the
same roots, a wonderful and charming forest of wild flowers, worthy
of the Palatines, consule dignæ. I can only tell you that this mixture
of art and reality is indescribable; it is at once a contest and a
harmony. Nature, who has a rival in Beethoven, finds also a rival in
Jean Goujon. The arabesques form tendrils and the tendrils form
arabesques. One does not know which to admire most, the living or
the sculptured leaf.
This ruin appears to be filled with a divine order.
It seems to me that this Palace, built by the fairies of the
Renaissance, is now in its natural state. All these marvellous
fantasies of free and savage art would be out of harmony in these
halls when treaties of peace or war were signed here, when grave
princes dreamed here, and when queens were married and German
emperors created here. Could these Vertumnuses, Pomonas, or
Ganymedes have understood anything about the ideas that came
into the heads of Frederick IV. or Frederick V., by the grace of God
Count Palatine of the Rhine, Vicar of the Holy Roman Empire,
Elector and Duke of Upper and Lower Bavaria? A grand seigneur
slept in this chamber beside a king’s daughter, under a ducal
baldaquin; now there is neither seigneur, king’s daughter, baldaquin,
nor even ceiling to this chamber; it is now the home of the bind-
weed, and the wild mint is its perfume. It is well. It is better thus. This
adorable sculpture was made to be kissed by the flowers and looked
upon by the stars....
 The night had fallen, the clouds were spread over the sky, and
the moon had mounted nearly to the zenith, while I was still sitting on
the same stone, gazing into the darkness which had gathered
around me and into the shadows which I had within me. Suddenly
the town-clock far below me sounded the hour; it was midnight: I
rose and descended. The road leading to Heidelberg passes the
ruins. At the moment when I arrived before them, the moon, veiled
by the diffused clouds and surrounded by an immense halo, threw a
weird light upon this magnificent mass of mouldering ruins....
The ruin, always open, is deserted at this hour. The idea of
entering it possessed me. The two stone giants, who guard the stone
court, allowed me to pass. I crossed the dark porch, upon which the
iron portcullis still hangs, and entered the court. The moon had
almost disappeared beneath the clouds. There was only a pallid light
in the sky.
Nothing is grander than that which has fallen. This ruin,
illuminated in such a way, at such an hour, was indescribably sad,
gentle, and majestic. I fancied that in the scarcely perceptible rustling
of the trees and foliage there was something grave and respectful. I
heard no footstep, no voice, no breath. In the court there was neither
light, nor shadow; a sort of dreamful twilight outlined everything and
veiled everything. The confused gaps and rifts allowed the feeble
rays of moonlight to penetrate the most remote corners; and in the
black depths of the inaccessible arches and corridors, I saw white
figures, slowly gliding.
It was the hour when the façades of old abandoned buildings are
no longer façades, but faces. I walked over the uneven pavement
without daring to make any noise, and I experienced between the
four walls of this enclosure that strange disquietude, that undefined
sentiment which the ancients called “the horror of the sacred woods.”
There is a kind of insurmountable terror in the sinister mingled with
the superb.
However, I climbed up the green and damp steps of the old
stairway without rails and entered the old roofless dwelling of Otho-
Heinrich. Perhaps you will laugh; but I assure you that to walk at
night through chambers which have been inhabited by people,
whose doors are dismantled, whose apartments each have their
peculiar signification, saying to yourself: “Here is the dining-room,
here is the bedroom, here is the alcove, here is the mantel-piece,”—
and to feel the grass under your feet and to see the sky above your
head, is terrifying. A room which has still the form of a room and
whose ceiling has been lifted off, as it were like the lid of a box,
becomes a mournful and nameless thing. It is not a house, it is not a
tomb. In a tomb you feel the soul of a man; in this place you feel his
shadow.
As soon as I passed the Knights’ Hall I stopped. Here there was
a singular noise, the more distinct because a sepulchral silence filled
the rest of the ruin. It was a weak, prolonged, strident rattle, mingled
at moments with a little, dry and rapid hammering, which at times
seemed to come from the depths of the darkness, from a far-away
copse, or the edifice itself; at times, from beneath my feet between
the rifts in the pavement. Whence came this noise? Of what
nocturnal creature was it the cry, or the knocking? I am not
acquainted with it, but as I listen to it, I cannot help thinking of that
hideous, legendary spinner who weaves rope for the gibbet.
However, nothing, nobody, not a living person is here. This hall,
like the rest of the Palace, is deserted. I struck the pavement with my
cane, the noise ceased, only to begin again a moment afterwards. I
knocked again, it ceased, then it began again. Yet I saw nothing but
a large frightened bat, which the blow of my cane on the stones had
scared from one of the sculptured corbels of the wall, and which
circled around my head in that funereal flight which seems to have
been made for the interior of ruined towers....
At the moment I descended the flight of stairs the moon shone
forth, large and brilliant, from a rift in the clouds; the Palace of
Frederick IV., with its double pediment, suddenly appeared,
magnificent and clear as daylight with its sixteen pale and formidable
giants; while, at my right, Otho’s façade, a black silhouette against
the luminous sky, allowed a few dazzling rays of moonlight to escape
through its twenty-four windows.
 I said clear as daylight—I am wrong. The moon upon ruins is
more than a light,—it is a harmony. It hides no detail, it exaggerates
no wounds, it throws a veil on broken objects and adds an
indescribable, misty aureole of majesty to ancient buildings. It is
better to see a palace, or an old cloister, at night than in the day. The
hard brilliancy of the sunlight is severe upon the ruins and intensifies
the sadness of the statues....
I went out of the Palace through the garden, and, descending, I
stopped once more for a moment on one of the lower terraces.
Behind me the ruin, hiding the moon, made, half down the slope, a
large mass of shadow, where in all directions were thrown out long,
dark lines, and long, luminous lines, which striped the vague and
misty background of the landscape. Below me lay drowsy
Heidelberg, stretched out at the bottom of the valley, the length of
the mountain; all the lights were out; all the doors were shut; below
Heidelberg I heard the murmur of the Neckar, which seemed to be
whispering to the hill and valley; and the thoughts which filled me all
the evening,—the nothingness of man in the Past, the infirmity of
man in the Present, the grandeur of Nature, and the eternity of God,
—came to me altogether, in a triple figure, whilst I descended with
slow steps into the darkness between this river awake and living, this
sleeping town, and this dead Palace.

Le Rhin (Paris, 1842).

 THE DUCAL PALACE.
JOHN RUSKIN.

THE charm which Venice still possesses, and which for the last fifty
years has rendered it the favourite haunt of all the painters of
picturesque subject, is owing to the effect of the palaces belonging to
the period we have now to examine, mingled with those of the
Renaissance.
The effect is produced in two different ways. The Renaissance
palaces are not more picturesque in themselves than the club-
houses of Pall Mall; but they become delightful by the contrast of
their severity and refinement with the rich and rude confusion of the
sea life beneath them, and of their white and solid masonry with the
green waves. Remove from beneath them the orange sails of the
fishing boats, the black gliding of the gondolas, the cumbered decks
and rough crews of the barges of traffic, and the fretfulness of the
green water along their foundations, and the Renaissance palaces
possess no more interest than those of London or Paris. But the
Gothic palaces are picturesque in themselves, and wield over us an
independent power. Sea and sky, and every other accessory might
be taken away from them, and still they would be beautiful and
strange. They are not less striking in the loneliest streets of Padua
and Vicenza (where many were built during the period of the
Venetian authority in those cities) than in the most crowded
thoroughfares of Venice itself; and if they could be transported into
the midst of London, they would still not altogether lose their power
over the feelings.
The best proof of this is in the perpetual attractiveness of all
pictures, however poor in skill, which have taken for their subject the
principal of these Gothic buildings, the Ducal Palace. In spite of all
architectural theories and teachings, the paintings of this building are
always felt to be delightful; we cannot be wearied by them, though
often sorely tried; but we are not put to the same trial in the case of
the palaces of the Renaissance. They are never drawn singly, or as
the principal subject, nor can they be. The building which faces the
Ducal Palace on the opposite side of the Piazzetta is celebrated
among architects, but it is not familiar to our eyes; it is painted only
incidentally, for the completion, not the subject, of a Venetian scene;
and even the Renaissance arcades of St. Mark’s Place, though
frequently painted, are always treated as a mere avenue to its
Byzantine church and colossal tower. And the Ducal Palace itself
owes the peculiar charm which we have hitherto felt, not so much to
its greater size as compared with other Gothic buildings, or nobler
design (for it never yet has been rightly drawn), as to its comparative
isolation. The other Gothic structures are as much injured by the
continual juxtaposition of the Renaissance palaces, as the latter are
aided by it; they exhaust their own life by breathing it into the
Renaissance coldness: but the Ducal Palace stands comparatively
alone, and fully expresses the Gothic power....
The Ducal Palace, which was the great work of Venice, was built
successively in the three styles. There was a Byzantine Ducal
Palace, a Gothic Ducal Palace, and a Renaissance Ducal Palace.
The second superseded the first totally; a few stones of it (if indeed
so much) are all that is left. But the third superseded the second in
part only, and the existing building is formed by the union of the two.
We shall review the history of each in succession.
1st. The Byzantine Palace.
The year of the death of Charlemagne, 813,—the Venetians
determined to make the island of Rialto the seat of the government
and capital of their state. Their Doge, Angelo or Agnello Participazio,
instantly took vigorous means for the enlargement of the small group
of buildings which were to be the nucleus of the future Venice. He
appointed persons to superintend the rising of the banks of sand, so
as to form more secure foundations, and to build wooden bridges
over the canals. For the offices of religion he built the Church of St.
Mark; and on, or near, the spot where the Ducal Palace now stands,
he built a palace for the administration of the government.
The history of the Ducal Palace therefore begins with the birth of
Venice, and to what remains of it, at this day, is entrusted the last
representation of her power....

THE DUCAL PALACE

In the year 1106, it was for the second time injured by fire, but
repaired before 1116, when it received another emperor Henry V. (of
Germany), and was again honoured by imperial praise. Between
1173 and the close of the century, it seems to have been again
repaired and much enlarged by the Doge Sebastian Ziani....
2nd. The Gothic Palace.
 The reader, doubtless, recollects that the important change in
the Venetian government which gave stability to the aristocratic
power took place about the year 1297, under the Doge Pietro
Gradenigo, a man thus characterized by Sansovino:—“A prompt and
prudent man, of unconquerable determination and great eloquence,
who laid, so to speak, the foundations of the eternity of this republic,
by the admirable regulations which he introduced into the
government.”...
We accordingly find it recorded by Sansovino, that “in 1301
another saloon was begun on the Rio del Palazzo under the Doge
Gradenigo, and finished in 1309, in which year the Grand Council
first sat in it.” In the first year, therefore, of the Fourteenth Century,
the Gothic Ducal Palace of Venice was begun; and as the Byzantine
Palace, was, in its foundation, coeval with that of the state, so the
Gothic Palace, was, in its foundation, coeval with that of the
aristocratic power. Considered as the principal representation of the
Venetian school of architecture, the Ducal Palace is the Parthenon of
Venice, and Gradenigo its Pericles....
Its decorations and fittings, however, were long in completion;
the paintings on the roof being only executed in 1400. They
represented the heavens covered with stars, this being, says
Sansovino, the bearings of the Doge Steno.... The Grand Council sat
in the finished chamber for the first time in 1423. In that year the
Gothic Ducal Palace of Venice was completed. It had taken, to build
it, the energies of the entire period which I have above described as
the central one of her life.
3rd. The Renaissance Palace.
I must go back a step or two, in order to be certain that the
reader understands clearly the state of the palace in 1423. The
works of addition or renovation had now been proceeding, at
intervals, during a space of a hundred and twenty-three years. Three
generations at least had been accustomed to witness the gradual
advancement of the form of the Ducal Palace into more stately
symmetry, and to contrast the works of sculpture and painting with
which it was decorated,—full of the life, knowledge, and hope of the
Fourteenth Century,—with the rude Byzantine chiselling of the
palace of the Doge Ziani. The magnificent fabric just completed, of
which the new Council Chamber was the nucleus, was now
habitually known in Venice as the “Palazzo Nuovo;” and the old
Byzantine edifice, now ruinous, and more manifest in its decay by its
contrast with the goodly stones of the building which had been raised
at its side, was of course known as the “Palazzo Vecchio.” That
fabric, however, still occupied the principal position in Venice. The
new Council Chamber had been erected by the side of it towards the
Sea; but there was not the wide quay in front, the Riva dei Schiavoni,
which now renders the Sea Façade as important as that of the
Piazzetta. There was only a narrow walk between the pillars and the
water; and the old palace of Ziani still faced the Piazzetta, and
interrupted, by its decrepitude, the magnificence of the square where
the nobles daily met. Every increase of the beauty of the new palace
rendered the discrepancy between it and the companion building
more painful; and then began to arise in the minds of all men a
vague idea of the necessity of destroying the old palace, and
completing the front of the Piazzetta with the same splendour as the
Sea Façade.... The Great Council Chamber was used for the first
time on the day when Foscari entered the Senate as Doge,—the 3rd
of April, 1423, ... and the following year, on the 27th of March, the
first hammer was lifted up against the old palace of Ziani.
That hammer stroke was the first act of the period properly
called the “Renaissance.” It was the knell of the architecture of
Venice,—and of Venice herself....
The whole work must have been completed towards the middle
of the Sixteenth Century.... But the palace was not long permitted to
remain in this finished form. Another terrific fire, commonly called the
great fire, burst out in 1574, and destroyed the inner fittings and all
the precious pictures of the Great Council Chamber, and of all the
upper rooms on the Sea Façade, and most of those on the Rio
Façade, leaving the building a mere shell, shaken and blasted by the
flames.... The repairs necessarily undertaken at this time were
however extensive, and interfere in many directions with the earlier
work of the palace: still the only serious alteration in its form was the
Welcome to our website – the ideal destination for book lovers and
knowledge seekers. With a mission to inspire endlessly, we offer a
vast collection of books, ranging from classic literary works to
specialized publications, self-development books, and children's
literature. Each book is a new journey of discovery, expanding
knowledge and enriching the soul of the reade

Our website is not just a platform for buying books, but a bridge
connecting readers to the timeless values of culture and wisdom. With
an elegant, user-friendly interface and an intelligent search system,
we are committed to providing a quick and convenient shopping
experience. Additionally, our special promotions and home delivery
services ensure that you save time and fully enjoy the joy of reading.

Let us accompany you on the journey of exploring knowledge and

personal growth!

ebookultra.com